mukul975 avatar

data-classification-skills

12 data classification skills: auto-discovery, PII detection, data inventory, labeling, lineage tracking, special category data

version
1.0.0
GitHub stars
340

Install

/plugin install data-classification-skills@privacy-data-protection-skills

Run it inside Claude Code after adding the marketplace with /plugin marketplace add mukul975/Privacy-Data-Protection-Skills.

Contents

Discovered from source

Skills 12

Sensitive Data Classification for AI/ML Training Datasets

Overview

AI and machine learning models trained on personal data raise distinct classification challenges. Training data may contain direct personal data, inferred special categories, proxy variables for protected characteristics, and data whose consent scope does not extend to model training. The EU AI Act (Regulation (EU) 2024/1689) imposes additional requirements for high-risk AI systems, including data governance obligations under Art. 10 that intersect with GDPR classification requirements. This skill provides a framework for classifying training data, detecting bias-relevant features, documenting data provenance, and verifying consent coverage.

GDPR and AI Act Intersection

GDPR Requirements for Training Data

GDPR ArticleApplication to AI Training
Art. 5(1)(b) — Purpose limitationTraining a model is a distinct processing purpose; if data was collected for customer service, using it for ML training requires a compatible purpose assessment or new lawful basis
Art. 5(1)(c) — Data minimisationTraining datasets must not include more personal data than necessary for the model objective
Art. 6 — Lawful basisModel training requires its own lawful basis; legitimate interests (Art. 6(1)(f)) is most common, but requires LIA documentation
Art. 9 — Special categoriesIf training data contains or enables inference of special category data, Art. 9(2) condition required
Art. 22 — Automated decision-makingIf the trained model makes decisions with legal or significant effects, additional safeguards apply
Art. 25 — Data protection by designClassification of training data is a by-design measure enabling appropriate technical protections
Art. 35 — DPIAHigh-risk AI processing (profiling, automated decision-making) requires DPIA

EU AI Act Art. 10 — Data Governance for High-Risk AI

The AI Act Art. 10 requires that training, validation, and testing datasets for high-risk AI systems:

  1. Are subject to appropriate data governance and management practices (Art. 10(2))
  2. Are relevant, sufficiently representative, and to the extent possible free of errors and complete (Art. 10(3))
  3. Take into account the specific geographical, contextual, behavioural, or functional setting within which the AI system is intended to be used (Art. 10(4))
  4. Where special category data processing is strictly necessary for bias detection and correction, this is permitted under Art. 10(5), subject to appropriate safeguards including pseudonymisation and GDPR compliance

Training Data Classification Framework

Classification Tier 1: Personal Data Content

ClassificationDescriptionExample
TRAINING_PII_DIRECTDataset contains direct identifiersCustomer names, email addresses in NLP training corpus
TRAINING_PII_INDIRECTDataset contains indirect identifiersCustomer IDs, transaction patterns enabling singling out
TRAINING_SPECIAL_CATDataset contains Art. 9 special category dataHealth records for medical diagnosis model
TRAINING_CRIMINALDataset contains Art. 10 criminal dataFraud transaction labels derived from criminal investigations
TRAINING_PSEUDONYMISEDPersonal data replaced with tokens but re-identification key existsPseudonymised customer data with mapping held by data team
TRAINING_ANONYMISEDData verified as anonymised per WP29 criteriaAggregated population statistics with k ≥ 10
TRAINING_SYNTHETICArtificially generated data with no real personal dataGAN-generated synthetic transaction data
TRAINING_NON_PERSONALNo personal data contentMarket price data, weather data, product specifications

Classification Tier 2: Bias and Proxy Detection

Even when a dataset does not directly contain Art. 9 special category data, it may contain proxy variables that correlate with protected characteristics:

Proxy VariableCorrelated Protected CharacteristicDetection Method
Postcode/ZIP codeRacial/ethnic origin, socioeconomic statusGeographic demographic analysis
First nameGender, ethnic origin, age cohortName demographics database lookup
Language preferenceEthnic origin, nationalityStatistical correlation analysis
Shopping patternsReligious belief (halal/kosher purchases), health statusPurchase category analysis
Web browsing historyPolitical opinions, sexual orientation, health statusTopic modelling on browsing categories
Employment gap patternsGender (maternity), disability, healthStatistical pattern analysis
Credit scoreRacial/ethnic origin (documented correlation in US/UK studies)Disparate impact analysis

Classification Tier 3: Consent and Purpose Scope

ClassificationDescriptionCompliance Requirement
CONSENT_COVERS_TRAININGOriginal consent explicitly covers AI/ML trainingDocument consent text and verify specificity
CONSENT_DOES_NOT_COVEROriginal consent did not anticipate ML trainingNew consent required or alternative lawful basis needed
LEGITIMATE_INTERESTML training relies on legitimate interests (Art. 6(1)(f))Documented LIA required
CONTRACT_PERFORMANCEML training is necessary for contract performanceNarrow scope — must be genuinely necessary
PUBLIC_DATAData sourced from publicly available sourcesStill requires lawful basis; public availability is not a lawful basis
RESEARCH_EXEMPTIONProcessing under Art. 89(1) research exemptionAppropriate safeguards including pseudonymisation required

Data Card Documentation

A data card is a structured document accompanying each training dataset, providing transparency about its contents, provenance, and limitations. Modelled on the "Datasheets for Datasets" framework (Gebru et al., 2021) and adapted for GDPR compliance.

Required Data Card Fields for Vanguard Financial Services

SectionFields
1. Dataset IdentityName, version, creation date, owner, purpose
2. Personal Data ClassificationTier 1 classification, data elements present, classification labels
3. Data SubjectsCategories of data subjects, volume, geographic scope
4. ProvenanceOriginal collection purpose, source systems, processing chain from collection to training set
5. Consent/Lawful BasisTier 3 classification, consent text reference or LIA reference, purpose compatibility assessment
6. Special Category AssessmentWhether Art. 9 data is present (direct or inferred), Art. 9(2) condition if applicable
7. Bias AssessmentProxy variables identified, disparate impact analysis results, demographic representation statistics
8. De-identificationTechnique applied (pseudonymisation, anonymisation, synthetic generation), assessment reference
9. RetentionTraining data retention period, model retention period, deletion schedule
10. Access ControlsWho can access the training data, who can access the model, audit logging
11. DPIA ReferenceDPIA document reference if applicable
12. LimitationsKnown biases, geographic limitations, temporal limitations, data quality issues

Bias Detection Methodology

Step 1: Demographic Representation Analysis

For each training dataset, calculate representation statistics:

  • What percentage of records come from each demographic group (to the extent known)?
  • Does the representation match the target population?
  • Are any groups under-represented by more than 20% relative to population?

Step 2: Proxy Variable Identification

Scan all features for proxy correlation with Art. 9 protected characteristics:

  • Calculate correlation coefficient between each feature and known protected characteristics (where available)
  • Flag features with |correlation| > 0.3 as potential proxies
  • Document all proxy variables in the data card

Step 3: Disparate Impact Analysis

For classification or scoring models:

  • Calculate model performance metrics by demographic group
  • Apply the 80% (four-fifths) rule: if the selection rate for any protected group is less than 80% of the rate for the most favoured group, disparate impact may exist
  • Document disparate impact analysis results in the data card

Step 4: AI Act Art. 10(5) Assessment

If bias detection requires processing special category data:

  • Document why processing is "strictly necessary" for bias detection and correction
  • Implement pseudonymisation of the special category data used for bias testing
  • Ensure GDPR Art. 9(2) condition is established (typically Art. 9(2)(g) substantial public interest or Art. 9(2)(j) research)
  • Process in a controlled environment with access restricted to the bias assessment team
  • Delete special category data after bias assessment is complete

Enforcement and Regulatory Precedents

  • Italian Garante — Clearview AI (2022): EUR 20 million fine for processing biometric data scraped from public sources for AI facial recognition training without lawful basis, consent, or transparency. Established that public availability does not provide lawful basis for AI training.
  • Italian Garante — ChatGPT/OpenAI (2023): Temporary ban and subsequent enforcement requiring OpenAI to establish lawful basis for training data collection, implement age verification, and provide opt-out mechanisms. Highlighted that AI training on personal data requires GDPR compliance throughout the data lifecycle.
  • CNIL — Enforcement Notice on AI Training Data (2024): CNIL published guidance sheets on AI training data requiring purpose limitation assessment, proportionality analysis, and specific measures when training data contains special category data.
  • ICO — Generative AI and Data Protection Consultation (2024): ICO's position that legitimate interests is the most likely lawful basis for AI training but requires documented LIA considering data subject reasonable expectations.

Integration Points

  • personal-data-test: Training data must first be classified as personal or non-personal
  • special-category-data: Art. 9 data in training sets requires heightened protections
  • pseudo-vs-anon-data: De-identification of training data must be validated
  • classification-policy: Training data classified under enterprise classification tiers
  • data-lineage-tracking: Full provenance from original collection to model deployment must be tracked

Automated PII Discovery and Classification

Overview

Automated data discovery tools scan structured and unstructured data repositories to identify, classify, and catalogue personal data across the enterprise. Manual data inventories cannot keep pace with the volume, velocity, and variety of modern data processing. Automated discovery provides continuous visibility into where personal data resides, how it flows, and whether it is classified and protected according to policy. This skill covers implementation patterns for four leading platforms — Microsoft Purview, BigID, OneTrust DataDiscovery, and AWS Macie — with focus on scanning configuration, accuracy optimisation, and integration with privacy compliance workflows.

Platform Comparison

CapabilityMicrosoft PurviewBigIDOneTrust DataDiscoveryAWS Macie
Structured data scanningSQL Server, Azure SQL, Synapse, Cosmos DB, Oracle, PostgreSQL, MySQL, Teradata100+ connectors including all major RDBMS, NoSQL, data warehouses200+ connectors, pre-built integrations with SaaS applicationsS3, DynamoDB, RDS (via Lambda)
Unstructured data scanningSharePoint, OneDrive, Exchange, Azure Blob, Azure Files, AWS S3, GCP StorageFile shares, email, SharePoint, cloud storage, Slack, Teams, ConfluenceFile shares, email, cloud storage, collaboration platformsS3 buckets (primary focus)
Classification method300+ built-in sensitive information types (SITs), trainable classifiers, exact data match (EDM), custom regexML-based NER, correlation analysis, pattern matching, custom classifiersPattern matching, NER, contextual analysis, custom rulesML-based pattern matching, custom data identifiers, managed data identifiers
GDPR-specific classifiersEU national ID formats, EU passport numbers, EU debit/credit card numbers, EU tax ID numbers per Member StateGDPR personal data taxonomy, Art. 9 special category detection, cross-regulation mappingPre-built GDPR data subject types, purpose mapping, lawful basis taggingEU personal data identifiers (limited — primarily financial and identity patterns)
Accuracy tuningConfidence levels (low/medium/high), custom keyword dictionaries, EDM for exact matching, document fingerprintingML model retraining, feedback loop, confidence thresholds, correlation rulesConfidence scoring, validation rules, exception managementCustom data identifiers with regex and keyword proximity, severity scoring
Deployment modelSaaS (Microsoft 365/Azure), hybrid with Purview governanceSaaS, on-premises, hybridSaaS, on-premises agentAWS-native SaaS
Pricing modelPer information protection unit (Azure), per Microsoft 365 licence tier (E5 includes advanced)Per data source connector, per TB scannedPer data source module, per connectorPer S3 bucket evaluated, per GB scanned

Implementation Pattern — Microsoft Purview

Architecture

Data Sources                    Microsoft Purview
┌─────────────┐                ┌───────────────────────────┐
│ Azure SQL   │──────scanner──►│ Data Map (metadata store)  │
│ SharePoint  │──────scanner──►│ Data Catalog (search/tag)  │
│ AWS S3      │──────scanner──►│ Data Estate Insights       │
│ On-prem SQL │──self-hosted──►│ Information Protection     │
│ Power BI    │──────scanner──►│ Data Loss Prevention (DLP) │
└─────────────┘                └───────────────────────────┘

Configuration Steps for Vanguard Financial Services

Step 1: Register Data Sources

  • Register all Azure SQL databases, SharePoint sites, Azure Blob Storage accounts
  • Install self-hosted integration runtime for on-premises SQL Server and file shares
  • Register AWS S3 buckets via cross-cloud connector
  • Register Power BI workspaces for report-level sensitivity scanning

Step 2: Configure Scanning Rules

  • Enable built-in sensitive information types (SITs) for all EU jurisdictions:
    • UK National Insurance Number (pattern: [A-CEGHJ-PR-TW-Z]{2}\d{6}[A-D])
    • EU Passport Numbers (per Member State format)
    • International Banking Account Number (IBAN)
    • EU Tax Identification Numbers
    • EU Health Insurance Card Number
  • Create custom SITs for Vanguard-specific identifiers:
    • Customer account number format: VFS-\d{10}
    • Employee ID format: EMP-[A-Z]{2}\d{6}
    • Investment portfolio reference: PF-\d{8}-[A-Z]{2}
  • Configure confidence levels: HIGH for direct identifiers, MEDIUM for indirect, LOW for contextual matches
  • Enable exact data match (EDM) for customer name verification against CRM database

Step 3: Set Scanning Schedule

  • Full scan: Monthly (first Saturday of each month, 02:00 UTC)
  • Incremental scan: Weekly (every Saturday, 02:00 UTC)
  • On-demand scan: Triggered by new data source registration or incident response
  • Estimated scan duration: 8 hours for full scan across 15 TB of data

Step 4: Configure Sensitivity Labels

  • Map Purview sensitivity labels to Vanguard classification tiers:
    • Public → Purview label: Public
    • Internal → Purview label: General
    • Confidential → Purview label: Confidential
    • Restricted → Purview label: Highly Confidential (auto-applied to Art. 9/Art. 10 data)
  • Enable auto-labelling policies: Apply labels based on detected SITs
  • Configure label inheritance: Child items inherit parent container label unless higher

Step 5: DLP Policy Integration

  • Create DLP policies triggered by sensitivity labels:
    • Restricted data: Block external sharing, require encryption, notify DPO
    • Confidential data: Warn on external sharing, audit log, require justification
    • Internal data: Audit log only
  • Configure endpoint DLP for Windows devices (prevent copy to USB, print, screenshot for Restricted content)

Accuracy Tuning for Purview

IssueTuning Approach
False positive: UK phone numbers flagged as National Insurance numbersIncrease minimum confidence to HIGH for NINO SIT; add negative keyword list ("phone", "tel", "fax", "mobile")
False positive: Internal reference numbers flagged as account numbersCreate EDM schema for actual customer accounts; custom SIT with proximity to customer-related keywords
False negative: Health data in free-text email bodiesEnable trainable classifier for health content; train on sample of 50+ positive examples from occupational health correspondence
False negative: Genetic identifiers in research datasetsCreate custom SIT for rs-number pattern (rs\d{4,12}), ICD-10 codes, and HUGO gene names

Implementation Pattern — BigID

Architecture

BigID uses a distributed scanning architecture with correlation-based discovery:

Data Sources              BigID Platform
┌─────────────┐          ┌──────────────────────────┐
│ Databases   │──scan───►│ Discovery Engine          │
│ File Shares │──scan───►│ Correlation Engine (ML)   │
│ Cloud       │──scan───►│ Classification Engine     │
│ SaaS Apps   │──API────►│ Catalog & Inventory       │
│ Email       │──scan───►│ Privacy Rights Automation │
└─────────────┘          └──────────────────────────┘

Key Differentiator: Correlation-Based Identity Resolution

BigID's ML-based correlation engine identifies personal data by correlating data elements across sources to build identity profiles. This approach detects personal data that pattern matching alone would miss — for example, a customer ID in one system linked to a name in another.

Configuration for Vanguard

  • Data source connections: Configure connectors for Salesforce CRM, Workday HR, ADP Payroll, SharePoint, Azure SQL, AWS S3, on-premises Oracle Data Warehouse
  • Correlation rules: Enable identity correlation across CRM customer ID, HR employee ID, and payroll employee number
  • Classification policies: Map BigID classifiers to GDPR categories (personal data, special category, criminal data, pseudonymised, anonymised)
  • Scanning schedule: Incremental scans daily at 01:00 UTC; full correlation rebuild weekly

Implementation Pattern — OneTrust DataDiscovery

Architecture

OneTrust integrates discovery with its broader privacy management platform:

Data Sources              OneTrust Platform
┌─────────────┐          ┌──────────────────────────┐
│ Cloud/SaaS  │──API────►│ DataDiscovery Module      │
│ Databases   │──agent──►│ Data Mapping (Art. 30)    │
│ File Shares │──agent──►│ Assessment Automation     │
│ Endpoints   │──agent──►│ Consent Management        │
└─────────────┘          │ DSAR Automation           │
                         └──────────────────────────┘

Key Differentiator: Privacy Program Integration

OneTrust's value proposition is tight integration between discovery results and privacy program management — discovered personal data feeds directly into Art. 30 records, DPIA assessments, and DSAR fulfilment workflows.

Configuration for Vanguard

  • Data source inventory: Use OneTrust's data mapping questionnaires to establish baseline, then deploy automated scanning to validate and enrich
  • Integration with RoPA: Map discovered data categories to Art. 30 processing activity records
  • DSAR integration: Configure discovery results to support automated data subject access request fulfilment (locate all data for a given identity across systems)
  • Vendor discovery: Scan for personal data shared with third-party processors; reconcile with vendor management records

Implementation Pattern — AWS Macie

Architecture

Macie is purpose-built for S3 data discovery within AWS:

AWS Environment
┌─────────────────────────────────────┐
│  S3 Buckets    ──scan──►  Macie     │
│                          │          │
│  EventBridge  ◄──alerts──┘          │
│  Security Hub ◄──findings──┘        │
│  CloudWatch   ◄──metrics──┘         │
└─────────────────────────────────────┘

Configuration for Vanguard

  • Bucket selection: Enable Macie for all S3 buckets in eu-west-1 (London) and eu-west-2 (Ireland) regions
  • Custom data identifiers: Create identifiers for Vanguard-specific patterns (customer account numbers, employee IDs, portfolio references)
  • Managed identifiers: Enable all EU-region managed data identifiers (UK NINO, IBAN, EU passport)
  • Job scheduling: Scheduled classification jobs weekly; event-driven scan on new object upload for high-sensitivity buckets
  • Findings integration: Route findings to AWS Security Hub for centralised monitoring; trigger Lambda function for automated label application

Scanning Schedule Best Practices

Scan TypeFrequencyDuration WindowTrigger
Full discovery scanMonthlyWeekend maintenance window (8-12 hours)Scheduled
Incremental scanWeeklyOff-peak hours (2-4 hours)Scheduled
New source onboarding scanOn registrationWithin 48 hours of source registrationEvent-driven
Post-incident scanAs neededImmediate (targeted scope)Incident response
Pre-DPIA scanBefore DPIA commencement1-2 weeks before DPIA startProject-triggered

Accuracy Metrics and Tuning

Key Performance Indicators

MetricTargetMeasurement Method
Precision (true positive rate)> 90%Sample 100 classified items monthly; verify classification accuracy
Recall (detection rate)> 85%Plant known PII test data in scan scope; measure detection rate
False positive rate< 10%Count items classified as personal data that are not
False negative rate< 15%Count personal data items missed by the scanner
Classification consistency> 95%Same data element classified consistently across repeat scans

Continuous Improvement Cycle

Month 1: Baseline scan → establish initial accuracy metrics
Month 2: Review false positives/negatives → tune rules and thresholds
Month 3: Re-scan → measure improvement
Month 4: Expand scope (new data sources) → re-baseline
Month 5: Review edge cases → create custom classifiers
Month 6: Accuracy audit by DPO → formal accuracy report
[Repeat cycle]

Integration Points

  • data-inventory-mapping: Automated discovery feeds Art. 30 data inventory
  • pii-in-unstructured: Specialized techniques for unstructured data augment platform capabilities
  • data-labeling-system: Discovery results drive automated label application
  • classification-policy: Discovery validates compliance with classification policy tiers
  • data-lineage-tracking: Discovery combined with lineage provides complete data visibility

Data Classification Policy Development

Overview

A data classification policy establishes the enterprise-wide framework for categorising data by sensitivity level and prescribing handling requirements per tier. The policy translates GDPR classification obligations (personal, special category, criminal) and information security requirements (confidentiality, integrity, availability) into practical, enforceable rules that govern how data is stored, transmitted, accessed, and disposed of throughout its lifecycle. This skill covers the design, implementation, and enforcement of a four-tier classification scheme aligned with GDPR requirements and ISO 27001 Annex A controls.

Classification Tier Structure

Four-Tier Model for Vanguard Financial Services

TierLabelDescriptionGDPR MappingISO 27001 Mapping
Tier 1: PublicPUBLICInformation approved for public release. Disclosure carries no risk to the organisation or individuals.Anonymised data (Recital 26), published reports, public filingsA.8.2 — Unclassified/Public
Tier 2: InternalINTERNALInformation for internal use only. Disclosure would cause minor inconvenience but no significant harm.Pseudonymised aggregate data, internal procedures, organisational chartsA.8.2 — Internal
Tier 3: ConfidentialCONFIDENTIALInformation whose disclosure would cause significant harm to individuals or the organisation.Personal data (Art. 4(1)), financial data, customer records, employee recordsA.8.2 — Confidential
Tier 4: RestrictedRESTRICTEDInformation whose disclosure would cause severe harm, regulatory sanction, or irreversible damage.Art. 9 special category data, Art. 10 criminal data, trade secrets, regulatory investigation dataA.8.2 — Restricted/Secret

Classification Decision Matrix

Data TypeTierRationale
Published annual reportPUBLICApproved for public disclosure
Internal org chartINTERNALNo individual harm from disclosure, but not intended for external distribution
Customer name and emailCONFIDENTIALPersonal data — disclosure would breach GDPR obligations
Customer account number and transactionsCONFIDENTIALPersonal data with financial sensitivity
Employee National Insurance NumberCONFIDENTIAL (borderline RESTRICTED)High-sensitivity direct identifier
Employee health recordsRESTRICTEDArt. 9 special category data
Biometric access templatesRESTRICTEDArt. 9 biometric data
Criminal background check resultsRESTRICTEDArt. 10 criminal data
AML investigation filesRESTRICTEDCriminal data + legal privilege
Board strategic plansRESTRICTEDCommercial sensitivity
Encryption keys and credentialsRESTRICTEDSecurity-critical

Handling Procedures Per Tier

Tier 1: PUBLIC

ControlRequirement
StorageNo restrictions; any approved platform
TransmissionNo restrictions; may be sent via any channel
Access controlNo restrictions; accessible to all
EncryptionNot required (but recommended for integrity)
RetentionPer document retention schedule
DisposalStandard deletion
LabellingLabel: "Public" or no label required

Tier 2: INTERNAL

ControlRequirement
StorageVanguard-managed systems only (no personal devices, no personal cloud)
TransmissionVanguard email or approved collaboration tools; not via personal email
Access controlAll Vanguard employees and approved contractors with valid account
EncryptionRequired in transit (TLS 1.2+); recommended at rest
RetentionPer document retention schedule
DisposalStandard secure deletion (file system delete)
LabellingLabel: "Internal" applied via header/footer or metadata

Tier 3: CONFIDENTIAL

ControlRequirement
StorageVanguard-managed systems with access controls; encrypted at rest (AES-256)
TransmissionEncrypted email (TLS 1.3 or S/MIME); approved SFTP; no unencrypted channels
Access controlRole-based access control (RBAC); business need-to-know; MFA required
EncryptionRequired at rest (AES-256) and in transit (TLS 1.3); field-level encryption for high-sensitivity fields
PrintingPermitted with secure print release; printed copies secured in locked cabinets
External sharingPermitted only with approved third parties under DPA/NDA; DPO approval for new recipients
RetentionPer data-specific retention schedule; automated enforcement
DisposalSecure overwrite (NIST 800-88 Clear); shredding for physical media
LabellingLabel: "Confidential" in document header/footer, email subject tag, and metadata
AuditAccess logged; quarterly review of access permissions
DLPDLP policies active: warn on external sharing, block sharing to personal email

Tier 4: RESTRICTED

ControlRequirement
StorageDesignated restricted-access systems only; encrypted at rest with customer-managed keys; separate network segment or HSM where applicable
TransmissionEnd-to-end encrypted channels only; no email without encryption; approved secure file transfer only
Access controlNamed individual access lists (not role-based); dual-person authorisation for bulk access; MFA required; privileged access management (PAM)
EncryptionAES-256 at rest with customer-managed keys; TLS 1.3 in transit; field-level encryption mandatory
PrintingProhibited unless specifically authorised by DPO; watermarked with user identity
External sharingProhibited unless DPO and Chief Privacy Officer jointly approve; encrypted transfer only
RetentionStrict retention with automated deletion; no extensions without DPO approval
DisposalNIST 800-88 Purge (cryptographic erasure or physical destruction); disposal certificate required
LabellingLabel: "Restricted" in document header/footer (red), email banner, and metadata; visual marking on screens
AuditAll access logged with user identity, timestamp, and purpose; monthly audit review by DPO
DLPDLP policies active: block all external sharing, block USB copy, block print, block screenshot; alert DPO on policy trigger
Incident responseAny unauthorised access or disclosure treated as data breach; immediate DPO notification

Policy Governance

Policy Document Structure

SectionContent
1. Purpose and ScopeDefines why classification is required and what data is covered
2. Classification TiersFour-tier definitions with examples
3. Roles and ResponsibilitiesData owner, data steward, DPO, IT Security, all employees
4. Classification ProceduresHow to classify new data, reclassification triggers
5. Handling RequirementsPer-tier controls matrix
6. Labelling RequirementsHow to apply labels (manual and automated)
7. Exception ProcessHow to request exceptions with risk acceptance
8. EnforcementMonitoring, audit, and consequences for violations
9. TrainingClassification awareness training requirements
10. Review ScheduleAnnual review; interim reviews on regulatory change

Roles and Responsibilities

RoleResponsibility
Data Owner (business unit head)Classify data for their business function; approve access requests; review classifications annually
Data Steward (departmental)Maintain classification labels; ensure handling compliance in their area; report classification issues
DPOOversee policy compliance; approve Restricted tier exceptions; conduct classification audits
IT SecurityImplement technical controls per tier; manage DLP policies; configure encryption; monitor audit logs
All EmployeesApply classification labels to data they create; handle data per tier requirements; report suspected misclassification
Privacy EngineeringConfigure automated classification tools; manage discovery platforms; tune detection accuracy

Exception Process

Exception Request
  │
  ├─► Step 1: Requestor documents the exception
  │     - What data? What tier? What control cannot be met?
  │     - Why is the exception needed?
  │     - What alternative controls are proposed?
  │     - Duration of exception requested
  │
  ├─► Step 2: Risk Assessment
  │     - Data Steward assesses the risk of the exception
  │     - For Confidential tier: Data Owner approves
  │     - For Restricted tier: DPO and CISO jointly approve
  │
  ├─► Step 3: Approval and Documentation
  │     - Exception recorded in exception register
  │     - Alternative controls documented and implemented
  │     - Expiry date set (maximum 12 months, renewable)
  │
  └─► Step 4: Review
        - Exception reviewed at expiry
        - Renewed only if original justification remains valid
        - Chronic exceptions trigger process improvement

Enforcement Mechanisms

MechanismImplementation
Automated labellingMicrosoft Purview auto-labelling applies Confidential/Restricted labels based on detected PII
DLP policiesMicrosoft Purview DLP blocks or warns on policy violations (external sharing, USB copy)
Access reviewsQuarterly certification of access permissions for Confidential; monthly for Restricted
Audit loggingAll access to Confidential and Restricted data logged and retained for 2 years
Classification auditsDPO conducts semi-annual audits sampling 100 items per tier for classification accuracy
Training complianceAnnual classification training required for all employees; completion tracked in LMS
Disciplinary policyViolations escalated per employee handbook: warning → formal warning → disciplinary action

Enforcement Precedents

  • ICO v Interserve Group (2022): GBP 4.4 million fine for inadequate security measures — the ICO noted that failure to classify data by sensitivity contributed to the inability to apply proportionate security controls, leading to a breach affecting 113,000 employees.
  • CNIL v Sergic (2019): EUR 400,000 fine for failing to implement adequate access controls on tenant personal data — absence of data classification meant all data was treated with the same (insufficient) controls.
  • AEPD v CaixaBank (2021): EUR 6 million fine — the DPA noted that inadequate data classification contributed to excessive data collection and retention, violating data minimisation and storage limitation principles.

Integration Points

  • personal-data-test: Classification policy tiers are assigned based on personal data classification results
  • special-category-data: Art. 9 data automatically assigned Restricted tier
  • criminal-data-handling: Art. 10 data automatically assigned Restricted tier
  • data-labeling-system: Labelling system implements the policy's labelling requirements
  • auto-data-discovery: Discovery tools validate that classification labels match detected data sensitivity

Criminal Conviction and Offence Data Handling — GDPR Art. 10

Overview

Article 10 of the GDPR establishes a separate regime for processing personal data relating to criminal convictions and offences, or related security measures. Unlike Art. 9 special category data which is subject to a general prohibition with listed exceptions, Art. 10 permits processing only under the control of official authority, or when authorised by EU or Member State law providing appropriate safeguards. The maintenance of a comprehensive register of criminal convictions is restricted to processing under the control of official authority. This skill provides a framework for identifying, classifying, and lawfully processing criminal data within enterprise contexts.

Legal Framework — Art. 10

Text of Article 10

"Processing of personal data relating to criminal convictions and offences or related security measures based on Article 6(1) shall be carried out only under the control of official authority or when the processing is authorised by Union or Member State law providing for appropriate safeguards for the rights and freedoms of data subjects. Any comprehensive register of criminal convictions shall be kept only under the control of official authority."

Three Key Requirements

RequirementDescriptionPractical Implication
Official authority controlProcessing must be supervised by or conducted under the authority of a public body with legal competencePrivate sector organisations generally cannot process criminal data unless authorised by specific national law
National law authorisationEU or Member State law must specifically authorise the processing and provide appropriate safeguardsControllers must identify the specific legal provision authorising each criminal data processing activity
Comprehensive register prohibitionNo private entity may maintain a comprehensive register of criminal convictionsAggregating criminal records across individuals for database purposes requires official authority status

Relationship to Art. 6 Lawful Basis

Art. 10 operates as an additional layer on top of Art. 6. A controller processing criminal data must satisfy BOTH:

  1. A lawful basis under Art. 6(1) — typically Art. 6(1)(c) (legal obligation) or Art. 6(1)(f) (legitimate interests, where national law permits)
  2. One of the Art. 10 conditions — official authority control OR national law authorisation with appropriate safeguards

Relationship to Art. 9

Criminal data is NOT listed in Art. 9(1) as a special category, but Member States may impose Art. 9-equivalent protections. Some data may fall under both Art. 9 and Art. 10 — for example, data about a criminal offence that also reveals racial profiling concerns or political activity.

Scope of Criminal Data

Data Elements Within Art. 10 Scope

CategoryExamplesNotes
Criminal convictionsConviction records, sentences, court judgmentsCore Art. 10 scope
Criminal offencesCharges, indictments, allegations of criminal conductIncludes unproven allegations
Related security measuresProbation orders, restraining orders, electronic monitoring conditionsPost-conviction measures
Criminal proceedingsArrest records, bail conditions, court appearance datesProcedural data
Acquittals and dismissalsRecords of charges dropped or acquittalsStill criminal data under Art. 10
Cautions and warningsPolice cautions, penalty notices for disorderOut-of-court disposals
Spent convictionsConvictions that are rehabilitated under national law (UK: Rehabilitation of Offenders Act 1974)May have additional protections under national law

Data Elements Outside Art. 10 Scope

DataReason
Civil litigation recordsArt. 10 covers criminal matters only
Regulatory enforcement actions (fines by supervisory authorities)Administrative, not criminal
Disciplinary proceedings (employment misconduct)Internal employment matter, not criminal
Credit defaults and county court judgmentsCivil debt matters
Self-reported general "good character" statementsNot specific criminal data

Boundary Cases at Vanguard Financial Services

ScenarioArt. 10 Applicable?Reasoning
DBS (Disclosure and Barring Service) check results for new hires in regulated rolesYESContains criminal conviction and caution data
FCA (Financial Conduct Authority) fitness and propriety checksYES — where criminal history is assessedFCA Senior Managers and Certification Regime requires criminal history disclosure
Anti-money laundering suspicious activity reports (SARs)BORDERLINE — YES when the SAR relates to suspected criminal activitySAR data may constitute data relating to criminal offences (suspected fraud, money laundering)
Internal investigation into suspected employee fraudYES — if the investigation relates to conduct that would constitute a criminal offenceEven where no charges are brought, investigation data relating to criminal offences falls under Art. 10
Sanctions screening resultsNO — unless a sanctions match relates to criminal conviction or offenceSanctions are typically administrative/regulatory measures

National Law Derogations

United Kingdom — Data Protection Act 2018

  • Schedule 1, Part 1: Conditions for processing criminal conviction data under Art. 10 — includes processing necessary for employment purposes, regulatory requirements, preventing or detecting unlawful acts, protecting the public
  • Schedule 1, Part 2: Substantial public interest conditions — includes safeguarding of children and vulnerable adults, regulatory requirements relating to fraud
  • Rehabilitation of Offenders Act 1974: Spent convictions must not be disclosed or used in most employment contexts. Exceptions exist for regulated activities (financial services, healthcare, childcare) listed in the Rehabilitation of Offenders Act 1974 (Exceptions) Order 1975
  • DBS Checks: Enhanced DBS checks available only for prescribed positions; basic DBS check available for any purpose with the individual's consent

Germany — BDSG §§ 26, 85

  • § 26 BDSG: Employee criminal data may be processed where necessary for the employment relationship and proportionate. Processing of criminal data of job applicants is permitted only where the offence is relevant to the specific role
  • Criminal Register Act (Bundeszentralregistergesetz, BZRG): Federal Central Criminal Register managed exclusively by the Federal Office of Justice. Private entities cannot maintain criminal registers

France — Code pénal, Art. 226-19; CNIL Authorisation

  • Criminal data processing by private entities requires specific CNIL authorisation or explicit legal basis
  • Bulletin No. 3 of the criminal record (casier judiciaire) may be requested by employers only for positions specified by law

Netherlands — UAVG Art. 32

  • Processing of criminal data by private entities permitted only where specifically authorised by law or with explicit consent of the data subject
  • Verklaring Omtrent het Gedrag (VOG/Certificate of Good Conduct) system managed by Justis (Screening Authority)

Classification Methodology

Step 1: Identify Criminal Data Elements

For each system, scan for data elements that may contain criminal data:

Detection PatternField ExamplesConfidence
Criminal record identifierscriminal_record_number, dbs_certificate_number, police_referenceHigh
Offence classificationsoffence_code, charge_description, conviction_typeHigh
Court and sentencing datacourt_name, sentence_type, sentence_duration, judge_nameHigh
Investigation referencesinvestigation_id, sar_reference, fraud_case_numberMedium
Background check resultsdbs_result, background_check_status, criminal_history_flagHigh
Free-text fields containing criminal terminologyAny field containing "convicted", "arrested", "charged", "offence", "sentence"Low — requires manual review

Step 2: Determine Art. 10 Applicability

For each identified element:

  1. Confirm the data relates to a criminal conviction, offence, or related security measure
  2. Distinguish from civil, administrative, or disciplinary matters
  3. For borderline cases (e.g., SARs, internal investigations), assess whether the underlying conduct would constitute a criminal offence

Step 3: Identify National Law Authorisation

For each Art. 10 data element:

  1. Identify the specific national law provision authorising the processing
  2. Confirm the law provides "appropriate safeguards" as required by Art. 10
  3. Document any additional national restrictions (e.g., Rehabilitation of Offenders Act spent conviction protections)

Step 4: Assess Comprehensive Register Risk

Review whether any system maintains what could constitute a "comprehensive register of criminal convictions":

  • A database aggregating criminal records across multiple individuals
  • A system that enables searching/querying criminal history across a population
  • Even partial aggregation may raise comprehensive register concerns if it covers a defined population comprehensively

Step 5: Apply Classification Label

  • CRIMINAL_ART10: Data within Art. 10 scope with identified national law authorisation
  • CRIMINAL_ART10_NO_AUTH: Data within Art. 10 scope WITHOUT identified national law authorisation (processing must cease until authorisation established)
  • CRIMINAL_SPENT: Spent conviction data subject to additional Rehabilitation of Offenders Act protections
  • NOT_CRIMINAL: Data reviewed and confirmed outside Art. 10 scope

Implementation at Vanguard Financial Services

Systems Processing Criminal Data

SystemData ElementsNational Law BasisPurpose
HR Recruitment PlatformDBS check results (basic and enhanced), criminal declaration formsUK DPA 2018 Sch.1 Part 1 para 1 (employment); Rehabilitation of Offenders Act 1974 (Exceptions) Order 1975Pre-employment screening for FCA-regulated roles
Compliance Case ManagementSAR data, internal investigation records, regulatory referral recordsUK Proceeds of Crime Act 2002 s.330-332 (SAR obligations); UK DPA 2018 Sch.1 Part 2 para 14 (preventing/detecting unlawful acts)AML compliance, fraud investigation
FCA Regulatory ReportingSenior Manager criminal history declarations, Approved Person criminal record disclosuresFinancial Services and Markets Act 2000 s.61 (fitness and propriety); FCA SUP 10CRegulatory fitness and propriety assessments
Third-Party Due DiligenceCriminal background check results for vendors and counterpartiesUK DPA 2018 Sch.1 Part 2 para 6 (regulatory requirements); Money Laundering Regulations 2017 reg.28Know Your Customer, vendor risk management

Safeguards Required

SafeguardImplementation
Access restrictionCriminal data accessible only to designated Compliance and HR personnel with specific role-based access. Named individual authorisation list maintained.
Purpose limitationCriminal data processed only for the specific purpose authorised by law. No secondary use for general HR analytics or performance management.
Retention limitationDBS check results retained for maximum 6 months after recruitment decision. SAR data retained per Proceeds of Crime Act retention schedules.
Data minimisationOnly the outcome of criminal checks recorded (clear/not clear + relevant details), not the full criminal record unless required by specific regulation.
Logging and auditAll access to criminal data logged with user identity, timestamp, and purpose. Quarterly audit of access logs by DPO.
Spent conviction handlingSystem flag for spent convictions. Spent convictions excluded from standard employment checks. Visible only for roles exempt under the Exceptions Order (FCA-regulated positions).

Enforcement Precedents

  • Metropolitan Police Service v Information Commissioner (UK First-Tier Tribunal, 2014): Tribunal addressed the scope of "criminal offence data" and confirmed that allegations of criminal conduct, even where unproven, constitute criminal data requiring Art. 10-equivalent protections
  • ICO Enforcement Notice — Experian Ltd (2020): ICO found Experian's processing of personal data in its marketing services business, which included data indicative of financial crime risk, lacked appropriate lawful basis and transparency
  • CNIL Decision SAN-2022-023 — Clearview AI: While primarily an Art. 9 biometric case, the decision also addressed Clearview's processing of data relating to criminal investigations and the requirement for official authority control when aggregating such data
  • Austrian DPA — CRIF GmbH (2021): Fined for maintaining a database of criminal-related financial data (insolvency linked to fraud convictions) without adequate national law authorisation under Art. 10

Integration Points

  • special-category-data: Criminal data may overlap with Art. 9 categories (e.g., political offences revealing political opinions)
  • data-inventory-mapping: Art. 30 RoPA must specifically identify Art. 10 data processing activities
  • classification-policy: Criminal data classified as "Restricted" tier with enhanced handling procedures
  • cross-jurisdiction-class: Criminal data protections vary significantly across jurisdictions — mapping required for multinational operations

Cross-Jurisdiction Data Classification Harmonisation

Overview

Multinational organisations must classify data under multiple regulatory frameworks simultaneously. GDPR special categories (Art. 9), CCPA sensitive personal information (Cal. Civ. Code §1798.140(ae)), HIPAA protected health information (45 CFR §160.103), and LGPD sensitive personal data (Art. 5-II) each define sensitive data differently with distinct requirements. This skill provides a cross-regulation mapping matrix enabling unified classification that satisfies all applicable jurisdictions, supporting Vanguard Financial Services' operations across the UK/EU, United States, and Brazil.

Jurisdictional Definitions Compared

GDPR Art. 9 — Special Categories (EU/UK)

CategoryDefinition Summary
Racial or ethnic originData revealing racial or ethnic background
Political opinionsPolitical views, party membership, voting
Religious or philosophical beliefsFaith, atheism, philosophical convictions
Trade union membershipUnion membership status or activities
Genetic dataInherited/acquired genetic characteristics (Art. 4(13))
Biometric dataPhysical/behavioural characteristics for unique identification (Art. 4(14))
Health dataPhysical/mental health, healthcare provision (Recital 35)
Sex life or sexual orientationSexual behaviour, preferences, orientation

Processing standard: General prohibition with ten exceptions (Art. 9(2)(a)-(j))

CCPA/CPRA — Sensitive Personal Information (California)

Cal. Civ. Code §1798.140(ae) defines sensitive personal information as:

CategoryDefinition Summary
Social Security, driver's licence, state ID, passportGovernment-issued identifiers
Account log-in with password/security credentialsFinancial account access data
Debit/credit card combined with security codePayment card data with authentication
Precise geolocationLocation within 1,850 feet (GPS-level)
Racial or ethnic originSame scope as GDPR
Religious or philosophical beliefsSame scope as GDPR
Union membershipSame scope as GDPR
Personal communications (mail, email, text)Content of non-public communications
Genetic dataDNA, genetic test results
Biometric dataPhysiological, biological, behavioural characteristics for identification
Health informationPhysical/mental health condition, diagnosis, treatment
Sex life or sexual orientationSexual activity, preferences

Consumer right: Right to limit use and disclosure of sensitive PI (§1798.121)

HIPAA — Protected Health Information (United States)

45 CFR §160.103 defines PHI as individually identifiable health information held or transmitted by a covered entity or business associate, in any form, that relates to:

ElementDescription
Physical or mental health conditionPast, present, or future
Provision of health careTreatment, services, payments
Payment for health careClaims, billing, insurance
IdentifiabilityInformation that identifies or could identify the individual

18 HIPAA Identifiers (45 CFR §164.514(b)): Names, geographic data (smaller than state), dates (except year), phone, fax, email, SSN, medical record number, health plan beneficiary number, account number, certificate/licence number, vehicle identifiers, device identifiers, web URLs, IP addresses, biometric identifiers, full-face photographs, any unique identifying number

Standard: Minimum necessary standard; use and disclosure limited to minimum necessary for purpose

LGPD Art. 5-II — Sensitive Personal Data (Brazil)

CategoryDefinition Summary
Racial or ethnic originSame scope as GDPR
Religious beliefReligious conviction or practice
Political opinionPolitical views or affiliations
Trade union membershipUnion affiliation
Religious, philosophical, or political organisationBroader than GDPR — includes organisational membership
Health dataPhysical/mental health
Sex lifeSexual activity
Genetic dataGenetic information
Biometric dataBiometric characteristics for identification

Processing standard: Consent must be specific and highlighted; other bases available under Art. 11(II)

Cross-Regulation Mapping Matrix

Data CategoryGDPR Art. 9CCPA §1798.140(ae)HIPAA PHILGPD Art. 5-IIUnified Tier
Racial/ethnic originSpecial categorySensitive PIN/A (unless health-related)SensitiveRESTRICTED
Political opinionsSpecial categoryN/AN/ASensitiveRESTRICTED
Religious beliefsSpecial categorySensitive PIN/ASensitiveRESTRICTED
Trade union membershipSpecial categorySensitive PIN/ASensitiveRESTRICTED
Genetic dataSpecial categorySensitive PIPHI (if health-related)SensitiveRESTRICTED
Biometric data (for identification)Special categorySensitive PIPHI (identifier)SensitiveRESTRICTED
Health dataSpecial categorySensitive PIPHISensitiveRESTRICTED
Sex life/orientationSpecial categorySensitive PIN/A (unless health-related)SensitiveRESTRICTED
Social Security NumberPersonal (direct identifier)Sensitive PIPHI (identifier)Personal (direct)RESTRICTED
Precise geolocationPersonal (may be special if reveals religion/health)Sensitive PIPHI (if health-related)PersonalCONFIDENTIAL+
Financial account with credentialsPersonal (financial)Sensitive PIN/APersonalRESTRICTED
Personal communications contentPersonalSensitive PIN/APersonalCONFIDENTIAL
Medical record numberPersonal (indirect)N/A (separate HIPAA regime)PHI (identifier)PersonalRESTRICTED (in healthcare context)
Health insurance IDPersonal (indirect)N/APHI (identifier)PersonalRESTRICTED (in healthcare context)

Harmonised Classification Approach

Principle: Apply the Highest Standard

When data falls under multiple jurisdictions, apply the most protective classification:

For each data element:
  1. Classify under GDPR (personal / special category / criminal)
  2. Classify under CCPA (personal information / sensitive PI)
  3. Classify under HIPAA (PHI / non-PHI) — if applicable
  4. Classify under LGPD (personal / sensitive) — if applicable
  5. Assign the unified tier = MAX(all jurisdictional classifications)

Unified Tier Assignment Rules

Jurisdictional ClassificationUnified Tier
Any regulation classifies as most-sensitiveRESTRICTED
GDPR special category OR CCPA sensitive PI OR LGPD sensitiveRESTRICTED
HIPAA PHI (any)RESTRICTED
GDPR personal data AND CCPA personal information (not sensitive)CONFIDENTIAL
Non-personal under all applicable regulationsPUBLIC or INTERNAL

Implementation for Vanguard Financial Services

Jurisdictional Scope

JurisdictionApplicable LawVanguard Operations
EU/EEAGDPREU customer base, Frankfurt data centre
United KingdomUK GDPR + DPA 2018London headquarters, UK customer base
California, USACCPA/CPRAUS operations, California residents' data
Other US StatesVarious (VCDPA, CPA, CTDPA, etc.)US operations
BrazilLGPDBrazilian customer base

Cross-Jurisdiction Handling Requirements

For data that falls under multiple jurisdictions:

ScenarioJurisdictionsHandling
UK customer health dataUK GDPR + potentially HIPAA if shared with US health insurerRESTRICTED; Art. 9(2) condition + HIPAA BAA if shared with US covered entity
California employee biometric data (fingerprint)CCPA + Illinois BIPA + GDPR (if EU data subject)RESTRICTED; CCPA consent + BIPA written consent + GDPR Art. 9(2)(a)
Brazilian customer financial dataLGPD + GDPR (if transferred from EU)CONFIDENTIAL; LGPD lawful basis + GDPR Chapter V transfer safeguards
Cross-border employee HR dataUK GDPR + CCPA + LGPDCONFIDENTIAL (RESTRICTED if contains health/diversity data); satisfy all three frameworks simultaneously

Regulatory Precedents

  • Meta Platforms Ireland (EDPB Binding Decision 01/2023): EUR 1.2 billion fine for EU-US data transfers — underscores the importance of jurisdictional mapping when data crosses borders
  • Sephora (CCPA Enforcement, 2022): USD 1.2 million settlement with California AG for failure to honour opt-out of sale of personal information — demonstrated CCPA-specific classification obligations
  • ANPD Brazil (2023): First LGPD administrative sanctions for inadequate security measures on sensitive personal data — confirmed LGPD enforcement trajectory

Integration Points

  • personal-data-test: GDPR personal data test is the foundation; cross-jurisdiction classification layers additional requirements
  • special-category-data: GDPR special categories map to sensitive data across other jurisdictions
  • classification-policy: Unified tier from cross-jurisdiction mapping feeds the enterprise classification policy
  • data-inventory-mapping: Data inventory must record applicable jurisdictions per data element

Data Inventory and Mapping for Classification

Overview

A data inventory is the foundational artefact for privacy compliance, providing a structured catalogue of all personal data processing activities across the organisation. GDPR Article 30 mandates that controllers maintain Records of Processing Activities (RoPA) containing specific fields including categories of personal data, purposes of processing, categories of recipients, and transfers to third countries. This skill focuses on building the data inventory that feeds into the RoPA, with emphasis on system-by-system discovery, data flow mapping, classification linkage, and legal basis assignment per data category.

Art. 30 Required Content for Controllers — Art. 30(1)

Required FieldDescriptionClassification Relevance
(a) Controller identityName and contact details of controller, joint controller, representative, DPOAdministrative — not classification-dependent
(b) Purposes of processingSpecific purpose for each processing activityPurpose determines applicable lawful basis and drives classification-to-handling mapping
(c) Categories of data subjectsTypes of individuals whose data is processed (customers, employees, applicants)Data subject category affects applicable protections (e.g., children, employees)
(d) Categories of personal dataTypes of personal data processed per activityDirect output of data classification — personal, special category, criminal
(e) Categories of recipientsWho receives the data, including processorsDetermines transfer and sharing controls
(f) Transfers to third countriesInternational transfers and safeguard mechanismsTriggers Chapter V compliance requirements
(g) Retention periodsEnvisaged time limits for erasure of different categoriesClassification tier drives retention policy
(h) Security measuresGeneral description of Art. 32 technical and organisational measuresClassification tier determines security control level

Data Inventory Construction Methodology

Phase 1: System Discovery (Weeks 1-3)

Step 1: Enterprise Application Inventory

Build a complete list of systems that process personal data:

Discovery MethodSourcesOutput
IT asset management systemCMDB, ServiceNow, LansweeperRegistered applications and servers
Procurement recordsVendor contracts, SaaS subscriptions, licence managementThird-party systems processing data
Network traffic analysisFirewall logs, proxy logs, DNS queriesUndocumented data flows to external services
Shadow IT discoveryCloud access security broker (CASB) logs, expense reports for SaaS toolsUnauthorised or unknown applications
Department interviewsBusiness unit leads, data stewardsDepartmental tools and local databases
Vanguard Financial Services — System Inventory Example
SystemTypeDepartmentData SubjectsClassification Priority
Salesforce CRMSaaSSales, Customer ServiceCustomers, prospectsTier 1
Workday HRSaaSHuman ResourcesEmployees, candidatesTier 1
ADP GlobalViewSaaSPayrollEmployeesTier 1
Oracle Data WarehouseOn-premisesData EngineeringAll subjectsTier 1
SharePoint OnlineSaaSAll departmentsAll subjectsTier 1
NICE ActimizeOn-premisesComplianceCustomers, counterpartiesTier 1
HID Global AccessOn-premisesFacilitiesEmployees, contractorsTier 2
Cority Occ HealthSaaSOccupational HealthEmployeesTier 2
Bloomberg TerminalSaaSTradingCounterpartiesTier 2
Microsoft TeamsSaaSAll departmentsEmployees, external contactsTier 2
Adobe AnalyticsSaaSMarketingWebsite visitorsTier 2
ServiceNow ITSMSaaSITEmployeesTier 3
Confluence WikiSaaSEngineeringEmployeesTier 3

Phase 2: Data Element Mapping (Weeks 3-6)

For each system, document every data element:

Step 2: Schema Extraction
  • Structured databases: Extract schema metadata (table names, column names, data types) via SQL metadata queries or data catalogue APIs
  • SaaS applications: Use API documentation to enumerate data objects and fields
  • Unstructured repositories: Catalogue document types and their typical personal data content
  • Automated discovery: Deploy scanning tools to identify PII in all repositories (see auto-data-discovery skill)
Step 3: Data Element Classification

For each data element, assign:

  • Personal data classification: PERSONAL_DIRECT, PERSONAL_INDIRECT, SPECIAL_CATEGORY, CRIMINAL_ART10, PSEUDONYMISED, ANONYMISED, NON_PERSONAL (see personal-data-test skill)
  • Classification tier: Public, Internal, Confidential, Restricted (see classification-policy skill)
  • Art. 6 lawful basis: The legal basis for processing this element for each stated purpose
  • Art. 9(2) condition: If special category, the specific processing condition

Phase 3: Data Flow Mapping (Weeks 6-9)

Step 4: Intra-System Data Flows

Document how data moves within the organisation:

Customer Onboarding Data Flow — Vanguard Financial Services

Customer ──[web form]──► Salesforce CRM
                              │
                    ┌─────────┼──────────┐
                    ▼         ▼          ▼
              Oracle DW   SharePoint   Email (welcome)
              (analytics) (documents)  (notifications)
                    │
                    ▼
              Adobe Analytics
              (web behaviour)
Step 5: External Data Flows (Third-Party Sharing)

Document all data sharing with external parties:

Sender SystemRecipientData CategoriesTransfer MechanismLegal Safeguard
Salesforce CRMSalesforce Inc. (processor)Customer personal dataSaaS hostingStandard Contractual Clauses (2021/914)
Workday HRWorkday Inc. (processor)Employee personal + special categorySaaS hostingEU-US Data Privacy Framework
ADP PayrollHMRC (statutory recipient)Employee financial dataSecure file transferLegal obligation Art. 6(1)(c)
NICE ActimizeNCA (statutory recipient)Customer AML data (Art. 10)SAR Online portalLegal obligation POCA 2002
Oracle DWDeloitte (auditor)Anonymised financial dataEncrypted file transferJoint controller agreement

Phase 4: Legal Basis and Purpose Mapping (Weeks 9-11)

Step 6: Purpose-Basis Matrix

For each processing activity, map data categories to purposes and legal bases:

Processing ActivityData CategoryPurposeArt. 6 BasisArt. 9(2) / Art. 10
Customer account managementName, email, address, account numberContract performanceArt. 6(1)(b)N/A
KYC identity verificationName, DOB, ID documents, addressLegal obligation (MLR 2017)Art. 6(1)(c)N/A
Employee payrollName, NINO, bank details, salaryContract performanceArt. 6(1)(b)N/A
Employee diversity monitoringEthnicity, sexual orientationEquality monitoringArt. 6(1)(a) / 6(1)(c)Art. 9(2)(a) / 9(2)(b)
Building access controlFingerprint templatePhysical securityArt. 6(1)(f)Art. 9(2)(a)
AML monitoringTransaction patterns, SAR dataLegal obligation (POCA 2002)Art. 6(1)(c)Art. 10 — DPA Sch.1 Pt.2 para 10
Website analyticsIP address, cookies, browsingLegitimate interestsArt. 6(1)(f)N/A
Marketing communicationsEmail, preferencesConsentArt. 6(1)(a)N/A

Phase 5: Documentation and Maintenance (Ongoing)

Step 7: Inventory Consolidation

Consolidate all discovery into the formal data inventory:

  • One record per data element per system
  • Cross-referenced to processing activities in the RoPA
  • Classification labels applied per element
  • Review dates set per classification tier
Step 8: Ongoing Maintenance
Maintenance ActivityFrequencyResponsible
Automated discovery scan reconciliationMonthlyPrivacy Engineering
New system onboarding inventoryPer new system deploymentSystem Owner + Privacy Analyst
Department data steward reviewQuarterlyData Stewards
Full inventory accuracy auditAnnuallyDPO
Classification reclassification reviewAnnually (borderline: 6-monthly)Privacy Analyst

Data Flow Diagram Standards

Notation Convention (Based on DFD Level 0/1)

SymbolMeaning
RectangleExternal entity (data subject, third party, regulator)
Rounded rectangleProcess (system or application)
Open rectangle (two parallel lines)Data store (database, file share)
ArrowData flow (labelled with data categories)
Dashed arrowOptional or conditional data flow
Red borderContains special category or criminal data
Blue borderInternational transfer

Enforcement Precedents

  • Österreichische Datenschutzbehörde (Austrian DPA) v Local Bakery Chain (2020): EUR 4,000 fine for failure to maintain Art. 30 records. The DPA noted that even small organisations must maintain processing records.
  • CNIL v Sergic (2019): EUR 400,000 fine, partly for inadequate data mapping — the controller could not demonstrate it knew where personal data was stored, leading to a data breach going undetected.
  • ICO v British Airways (2020): GBP 20 million fine, where inadequate data mapping contributed to delayed breach detection — the controller lacked visibility into personal data flows through its web infrastructure.

Integration Points

  • auto-data-discovery: Automated scanning feeds data elements into the inventory
  • personal-data-test: Classification decisions applied per data element in the inventory
  • special-category-data: Art. 9 data flagged and mapped to processing conditions
  • criminal-data-handling: Art. 10 data identified and mapped to national law authorisation
  • data-lineage-tracking: Lineage graphs extend the static inventory with dynamic flow information

Data Classification Labels and Tagging System

Overview

Data classification labels are the operational mechanism through which classification policy is enforced across the enterprise. Labels attach classification metadata to data assets — documents, emails, database records, and cloud resources — enabling automated enforcement of handling requirements through DLP policies, access controls, and encryption. This skill covers the design and implementation of a labelling system using Microsoft Purview Information Protection as the primary platform, with architecture patterns for automated labelling, user-applied labelling, label inheritance, and cross-platform propagation.

Label Taxonomy

Vanguard Financial Services Label Hierarchy

Vanguard Classification Labels
├── Public
│   └── (no sub-labels)
├── Internal
│   └── Internal - Project Confidential
├── Confidential
│   ├── Confidential - Customer Data
│   ├── Confidential - Employee Data
│   ├── Confidential - Financial Data
│   └── Confidential - Legal
└── Restricted
    ├── Restricted - Special Category (Art. 9)
    ├── Restricted - Criminal Data (Art. 10)
    ├── Restricted - AML Investigation
    └── Restricted - Board & Strategy

Label Properties

LabelColourVisual MarkingEncryptionDLP PolicyAuto-Apply
PublicGreenFooter: "Vanguard Financial Services — Public"NoneNoneNo
InternalBlueFooter: "Vanguard Financial Services — Internal Use Only"OptionalWarn on externalNo
ConfidentialAmberHeader + Footer: "CONFIDENTIAL"Azure RMS (AES-256)Warn + audit external; block personal emailYes (when PII detected with >85% confidence)
RestrictedRedHeader + Footer: "RESTRICTED" with red background; watermark on printAzure RMS (AES-256, double key encryption)Block all external; block USB/print; alert DPOYes (when Art. 9/Art. 10 data detected with >85% confidence)

Microsoft Purview Implementation Architecture

Component Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    Microsoft Purview                             │
│                                                                  │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────────┐  │
│  │ Sensitivity   │  │ Auto-labelling│  │ DLP Policies          │  │
│  │ Labels        │  │ Policies      │  │ (endpoint, email,     │  │
│  │ (definitions) │  │ (rules)       │  │  SharePoint, Teams)   │  │
│  └──────┬───────┘  └──────┬───────┘  └──────────┬───────────┘  │
│         │                  │                      │              │
│         ▼                  ▼                      ▼              │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │              Unified Label Application Engine             │   │
│  └──────────────────────────────────────────────────────────┘   │
│         │                  │                      │              │
└─────────┼──────────────────┼──────────────────────┼──────────────┘
          ▼                  ▼                      ▼
   ┌─────────────┐  ┌──────────────┐  ┌──────────────────────┐
   │ Office Apps  │  │ SharePoint/  │  │ Endpoint DLP         │
   │ (Word, Excel,│  │ OneDrive     │  │ (Windows devices)    │
   │  Outlook)    │  │              │  │                      │
   └─────────────┘  └──────────────┘  └──────────────────────┘

Auto-Labelling Configuration

Policy 1: Confidential — Customer PII Detection
SettingValue
Policy NameVFS-AutoLabel-Confidential-CustomerPII
ScopeAll SharePoint sites, OneDrive accounts, Exchange mailboxes
ConditionsContent contains ANY of: UK National Insurance Number (HIGH confidence), IBAN (HIGH), Vanguard Account Number (HIGH), Credit Card (HIGH + Luhn validated)
Minimum count1 instance of any SIT at HIGH confidence
Label appliedConfidential - Customer Data
Priority2 (overridden by Restricted auto-label)
Policy 2: Restricted — Special Category Detection
SettingValue
Policy NameVFS-AutoLabel-Restricted-SpecialCategory
ScopeAll SharePoint sites, OneDrive accounts, Exchange mailboxes
ConditionsContent contains ANY of: ICD-10 codes (MEDIUM+), health terminology trainable classifier (HIGH), biometric template format detection, genetic marker patterns
Minimum count1 instance at MEDIUM confidence or above
Label appliedRestricted - Special Category (Art. 9)
Priority1 (highest priority — overrides all other auto-labels)
Policy 3: Restricted — Criminal Data Detection
SettingValue
Policy NameVFS-AutoLabel-Restricted-CriminalData
ScopeHR SharePoint sites, Compliance SharePoint sites
ConditionsContent contains: DBS reference patterns, criminal conviction terminology, SAR reference numbers
Label appliedRestricted - Criminal Data (Art. 10)
Priority1

Label Inheritance Rules

RuleDescriptionImplementation
Container inheritanceItems in a labelled SharePoint site inherit the site's label as minimumSite sensitivity label propagates to new items; existing items retain higher label
Email attachment inheritanceAttachments inherit the email's label if attachment label is lowerOutlook plugin checks attachment label vs email label on send
Parent-child inheritanceChild documents inherit parent folder label as minimumSharePoint library policy; items cannot be labelled below folder label
No downgrade without approvalUsers cannot remove or downgrade labels without justificationLabel policy: require justification text for downgrade; audit log entry
Highest label winsWhen documents are merged or combined, the highest label appliesUser training + DLP monitoring for combined documents

User-Applied Labelling

Labelling Responsibilities

ScenarioWho LabelsHow
New document creationAuthorSelect label from Office ribbon (Word, Excel, PowerPoint)
New email compositionSenderSelect label from Outlook toolbar; mandatory before sending external
File upload to SharePointUploaderLabel prompt on upload if no label detected
Data export from systemExporterLabel selection required before export completes
Physical document printingPrinterClassification header/footer printed automatically; user selects tier if not auto-labelled

Mandatory Labelling Policy

SettingValue
Require label on documentsYes — all Word, Excel, PowerPoint documents must have a label before save
Require label on emailsYes — for emails to external recipients; recommended for internal
Default labelInternal (applied if user does not select; user can override up or down)
Justification for downgradeRequired — user must enter text justification; logged in audit
Justification for removalRequired — DPO-approved exception only

DLP Integration

DLP Policy Matrix

LabelExternal EmailUSB/RemovablePrintScreenshotCloud Upload
PublicAllowAllowAllowAllowAllow
InternalWarnWarnAllowAllowBlock (non-approved cloud)
ConfidentialWarn + auditBlockSecure printAllow (watermarked)Block
RestrictedBlockBlockBlock (unless DPO approved)BlockBlock

DLP Alert Routing

Alert SeverityLabel TriggerRouting
LowInternal label — external email warn overriddenSecurity team email
MediumConfidential label — external sharing attemptedSecurity team + Data Owner
HighRestricted label — any policy triggerDPO + CISO + immediate investigation
CriticalRestricted label — data exfiltration indicatorsDPO + CISO + Incident Response Team + 15-minute SLA

Label Propagation Across Platforms

Supported Platforms

PlatformLabel MethodPropagation
Microsoft 365 (Word, Excel, PowerPoint)Native sensitivity label in file metadataFull support — label travels with file
Outlook / ExchangeEmail header X-MS-Exchange-Organization-ClassificationLabel persists in message store and forwarded copies
SharePoint OnlineDocument library metadata + file metadataDual storage — library and file level
OneDrive for BusinessFile metadataSame as SharePoint
TeamsChannel/chat message metadataLimited — file attachments inherit, message labels in compliance
PDF exportVisual marking (header/footer) + XMP metadataVisual marking persists; metadata depends on PDF viewer
Azure SQL DatabaseColumn-level sensitivity classification (sys.sensitivity_classifications)Native Azure SQL feature; integrates with Purview
AWS S3Object tags (key: classification, value: tier)AWS-native tagging; read by Macie and IAM policies
On-premises file sharesFile metadata (NTFS ADS) via AIP Unified Labelling clientRequires AIP client installed on endpoints

Enforcement Precedents

  • ICO v Interserve Group (2022): GBP 4.4 million — failure to implement adequate data classification and labelling contributed to staff not recognising the sensitivity of data compromised in a breach
  • CNIL v Free Mobile (2022): EUR 300,000 — customer personal data stored without classification or access controls; staff could access all customer data regardless of role or need

Integration Points

  • classification-policy: Labelling system is the technical implementation of the classification policy
  • auto-data-discovery: Discovery results trigger auto-labelling for newly detected PII
  • pii-in-unstructured: PII detection in documents drives label recommendation or auto-application
  • data-inventory-mapping: Labels feed the data inventory with current classification status per asset
  • data-lineage-tracking: Labels propagate through data lineage — transformed data inherits source label

Data Lineage Tracking

Purpose

Map and document the complete lifecycle of personal data through organizational systems, from initial collection point through all transformations, transfers, and storage locations to final deletion or anonymization.

Prerequisites

  • Data inventory or asset register identifying systems processing personal data
  • Network architecture documentation showing system interconnections
  • Access to system metadata, ETL configurations, and API documentation
  • Records of Processing Activities (RoPA) under GDPR Article 30

Workflow

Step 1: Define Lineage Scope

Determine scope boundaries for the lineage mapping exercise:

  1. Data categories: Identify which personal data categories to trace (identity data, contact data, financial data, behavioral data, special category data under Art. 9)
  2. System boundaries: Define which systems are in-scope (production databases, data warehouses, analytics platforms, third-party SaaS, backup systems)
  3. Temporal scope: Determine whether to map current-state lineage only or include historical data flows
  4. Legal basis mapping: Link each processing activity in the lineage to its GDPR Article 6 lawful basis

Step 2: Identify Data Sources (Collection Points)

Document every point where personal data enters the organization:

  1. Direct collection: Web forms, mobile apps, point-of-sale terminals, customer service interactions, paper forms digitized via scanning
  2. Indirect collection: Third-party data providers, publicly available sources, data brokers, partner organizations
  3. Derived data: Data generated through processing (risk scores, customer segments, behavioral profiles)
  4. Inferred data: Data inferred from other data points (creditworthiness, health predictions, preferences)

For each source, record:

  • Source identifier and type
  • Data categories collected (referencing Art. 30(1)(c) categories)
  • Legal basis under Art. 6(1) and, if applicable, Art. 9(2)
  • Information provided to data subjects per Art. 13 or Art. 14
  • Volume and frequency of collection

Step 3: Map Data Transformations

Document every transformation applied to personal data:

  1. ETL processes: Extract-Transform-Load pipelines moving data between systems
  2. Aggregation: Grouping individual records into summary statistics
  3. Pseudonymization: Replacing identifiers with tokens per Art. 4(5) and Recital 26
  4. Anonymization: Irreversible de-identification per WP29 Opinion 05/2014
  5. Enrichment: Combining data from multiple sources to create enriched profiles
  6. Format conversion: Changing data formats (CSV to JSON, database migration)

For each transformation, record:

  • Input data categories and source system
  • Transformation logic description
  • Output data categories and destination system
  • Whether transformation changes the identifiability of data subjects
  • Retention period at destination per Art. 5(1)(e) storage limitation

Step 4: Document Data Flows and Transfers

Map all movements of personal data between systems and parties:

  1. Internal flows: Between departments, systems, databases within the organization
  2. Processor transfers: To data processors under Art. 28 agreements
  3. Third-country transfers: Cross-border transfers requiring Art. 44-49 safeguards
  4. Third-party disclosures: To independent controllers (regulators, partners, law enforcement)

For each flow, record:

  • Source and destination system/entity
  • Transfer mechanism (API, file transfer, database replication, manual export)
  • Legal safeguard for international transfers (SCCs, BCRs, adequacy decision)
  • Encryption in transit and at rest
  • Frequency and volume

Step 5: Map Data Storage and Retention

Document where personal data resides at each stage:

  1. Primary storage: Production databases, CRM systems, HRIS
  2. Secondary storage: Data warehouses, analytics databases, reporting systems
  3. Archival storage: Long-term archives, cold storage, compliance archives
  4. Backup storage: Disaster recovery systems, backup tapes, cloud backup
  5. Temporary storage: Caches, message queues, log files, session storage

For each storage location, record:

  • Storage technology and location (on-premises, cloud region)
  • Retention period and legal basis for retention
  • Access controls and encryption
  • Deletion or anonymization mechanism at end of retention period

Step 6: Implement Automated Lineage Discovery

Deploy tooling to automate lineage tracking:

  1. Database-level lineage: Query log analysis, column-level lineage from SQL parsing
  2. Application-level lineage: API call tracing, service mesh observability
  3. Pipeline-level lineage: ETL tool metadata (Apache Airflow lineage backend, dbt documentation)
  4. Infrastructure-level lineage: Network flow logs, data lake audit trails

Use the scripts/process.py helper to parse system metadata and generate lineage graphs.

Step 7: Integrate with RoPA and Compliance

Link lineage data to GDPR compliance documentation:

  1. Art. 30 RoPA: Each lineage path should map to a processing activity in the RoPA
  2. DPIA triggers: Flag lineage paths involving Art. 35(3) processing (large-scale profiling, systematic monitoring, special categories)
  3. Data subject rights: Use lineage to locate all data for DSAR responses (Art. 15 access, Art. 17 erasure, Art. 20 portability)
  4. Breach impact scoping: Use lineage to determine affected data subjects and categories during incident response per Art. 33(3)

Step 8: Maintain and Validate

Establish ongoing lineage maintenance:

  1. Change management: Update lineage when new systems, data flows, or processing activities are introduced
  2. Periodic validation: Quarterly review to verify lineage accuracy against actual system behavior
  3. Stakeholder review: Annual sign-off from data owners, system architects, and DPO
  4. Completeness check: Cross-reference lineage against data inventory and RoPA to identify gaps

Verification

  • All personal data categories have documented source-to-sink lineage paths
  • Each lineage node includes legal basis, retention period, and access controls
  • Transformation steps document identifiability impact
  • Cross-border transfers identified with Art. 44-49 safeguards documented
  • Lineage integrates with Art. 30 RoPA entries
  • Automated discovery covers at least database and pipeline layers
  • Quarterly validation schedule established with responsible owners assigned

Personal Data Classification Test — GDPR Art. 4(1)

Overview

Article 4(1) of the GDPR defines personal data as "any information relating to an identified or identifiable natural person ('data subject')." This definition is deliberately broad and technology-neutral. The European Court of Justice in Breyer v Bundesrepublik Deutschland (C-582/14, 19 October 2016) confirmed that even dynamic IP addresses can constitute personal data when the controller has legal means to obtain additional information enabling identification. This skill provides a systematic decision framework for classifying data elements as personal, non-personal, or borderline requiring contextual assessment.

Legal Foundation

Art. 4(1) — Four Constituent Elements

Personal data exists when ALL four elements are satisfied:

ElementDefinitionAssessment Criteria
Any informationNo restriction on nature, content, or format of informationIncludes objective facts (age, blood type) and subjective assessments (credit rating, performance review). Covers all formats: text, image, audio, biometric, metadata, behavioural
Relating toInformation must have a content, purpose, or result link to the individualContent link: information is about the person. Purpose link: information is used to evaluate or influence the person. Result link: processing has an impact on the person's rights or interests
Identified or identifiableThe person is or can be distinguished from all other personsIdentified: directly singled out. Identifiable: can be singled out by using additional data, taking into account all means reasonably likely to be used
Natural personLiving human being, not legal entities or deceased personsExcludes companies, government bodies, associations. Member State law may extend protections to deceased persons (e.g., Italy extends to 20 years post-mortem)

Recital 26 — "Reasonably Likely" Test for Identifiability

To determine whether a natural person is identifiable, account should be taken of all the means reasonably likely to be used, such as singling out, either by the controller or by another person to identify the natural person directly or indirectly. The assessment must consider:

  • All objective factors: cost of identification, time required, available technology at the time of processing, and technological developments anticipated during the retention period
  • All means reasonably likely: not limited to means the controller currently possesses — includes means any third party might reasonably employ
  • Dynamic assessment: what is not identifiable today may become identifiable as technology evolves or datasets are linked

The Breyer Ruling — CJEU C-582/14

Facts

Patrick Breyer challenged the German Federal Government's practice of storing dynamic IP addresses of visitors to government websites. Germany argued that dynamic IP addresses were not personal data because the website operator could not identify visitors without additional data held by the internet service provider (ISP).

Holding

The CJEU ruled that dynamic IP addresses constitute personal data for the website operator when:

  1. The operator has legal means available to obtain additional identifying information from a third party (the ISP)
  2. Identification does not require disproportionate effort in terms of time, cost, or labour
  3. The means are reasonably likely to be used — not merely theoretical

Key Legal Principles Established

  • Relative approach to identifiability: Personal data status is assessed relative to each controller's circumstances, not in the abstract
  • Legal means suffice: The controller need not currently possess the identifying data; having legal channels to obtain it is sufficient
  • Third-party knowledge counts: Information held by third parties must be considered if the controller has lawful means of access
  • Broad interpretation: The CJEU confirmed the GDPR's (then Directive 95/46/EC's) intent to apply broadly to protect fundamental rights

Practical Impact

After Breyer, the following are presumptively personal data for most controllers:

  • Dynamic and static IP addresses
  • Device fingerprints and browser fingerprints
  • Cookie identifiers and advertising IDs
  • MAC addresses when combined with network access logs
  • Pseudonymised datasets where re-identification keys exist or are obtainable

Decision Tree for Personal Data Classification

Stage 1: Is the Data About a Natural Person?

Data Element
    │
    ├── About a living natural person? ──► YES → Go to Stage 2
    │
    ├── About a deceased person? ──► Check Member State law (may still be protected)
    │
    ├── About a legal entity only? ──► NOT personal data under GDPR
    │       (but may contain personal data of individuals within,
    │        e.g., sole trader name = personal data)
    │
    └── About an anonymous aggregate? ──► Go to Stage 3 (verify truly anonymous)

Stage 2: Can the Person Be Identified or Is the Person Identifiable?

Data relates to a natural person
    │
    ├── Person is DIRECTLY identified?
    │   (name, photograph, unique ID number)
    │   ──► PERSONAL DATA
    │
    ├── Person is INDIRECTLY identifiable?
    │   (combination of data points enables singling out)
    │   ──► Apply Recital 26 "reasonably likely" test → Stage 2a
    │
    └── Person cannot be identified by any means reasonably likely?
        ──► NOT personal data (but document the assessment)

Stage 2a: Recital 26 Reasonably Likely Assessment

Indirect identifiers present
    │
    ├── Does the controller hold additional data enabling identification?
    │   ──► YES → PERSONAL DATA
    │
    ├── Does a third party hold such data, and does the controller
    │   have legal means to access it? (Breyer test)
    │   ──► YES → PERSONAL DATA
    │
    ├── Could publicly available data be combined to identify?
    │   (social media, public registers, news articles)
    │   ──► YES → PERSONAL DATA
    │
    ├── Is re-identification feasible considering:
    │   - Cost vs. value of identification
    │   - Time required vs. retention period
    │   - Current and foreseeable technology
    │   ──► YES → PERSONAL DATA
    │
    └── Identification requires disproportionate effort with no
        reasonable motivation?
        ──► NOT personal data (document reasoning)

Stage 3: Anonymisation Verification

Data claimed to be anonymous/aggregated
    │
    ├── Can any individual be singled out from the dataset?
    │   ──► YES → PERSONAL DATA (pseudonymised, not anonymised)
    │
    ├── Can records be linked to form a profile of an individual?
    │   ──► YES → PERSONAL DATA
    │
    ├── Can information be inferred about a specific individual?
    │   ──► YES → PERSONAL DATA
    │
    └── Passes all three tests (singling out, linkability, inference)?
        ──► Anonymised data — NOT personal data
        (Apply WP29 Opinion 05/2014 framework)

Classification Categories with Examples

Category A: Clear Personal Data (Always Personal)

Data ElementReason
Full nameDirect identifier
National ID number (SSN, Aadhaar, BSN)Unique direct identifier
Email address (personal)Directly identifies in most contexts
Photograph of a faceDirect visual identifier (also biometric if processed for identification)
Biometric data (fingerprint, iris scan)Unique to individual, Art. 9 special category when used for identification
Genetic dataUnique biological identifier, Art. 9 special category
Health records with patient nameDirect identifier plus Art. 9 special category
Home address with nameDirect identifier with location

Category B: Contextually Personal Data (Requires Assessment)

Data ElementWhen PersonalWhen Not Personal
Dynamic IP addressWhen controller has legal means to obtain subscriber info from ISP (Breyer)When controller has no means and no motivation to identify (rare)
Cookie identifierWhen linked to browsing profile that enables singling outWhen session-only cookie with no profile building
Device fingerprintWhen used to track across sites/sessionsWhen used only for aggregate device statistics with k-anonymity
Employee ID numberWhen linked to HR records by same controllerWhen used in anonymised survey with no re-identification key
Location data (GPS coordinates)When tracking individual movement patternsWhen aggregated to postcode-level with >1000 individuals per cell
Purchase historyWhen linked to customer accountWhen stripped of all identifiers and aggregated by product category
Vehicle registration numberWhen plate-to-owner lookup is legally availableNot applicable — plate lookup is available in most jurisdictions, so nearly always personal

Category C: Typically Not Personal Data

Data ElementCondition for Non-Personal Status
Weather dataGeneral environmental data not relating to individuals
Stock pricesCorporate financial data
Machine sensor readingsEquipment telemetry with no operator identification
Aggregated census statisticsPublished statistical tables with adequate anonymisation
Chemical compound propertiesScientific data about substances
Company financial statementsLegal entity data (but may contain director names)

Borderline Cases — Detailed Analysis

Case 1: Pseudonymised Data

Pseudonymised data remains personal data under GDPR (Recital 26, Art. 4(5)). The existence of a re-identification key — even if held by a separate entity — means the data relates to an identifiable person. Pseudonymisation is a security measure, not an anonymisation technique.

Vanguard Financial Services Application: Customer transaction records where account numbers are replaced with random tokens. The mapping table is held by a separate internal department with access controls. These remain personal data because:

  • Vanguard holds the re-identification key internally
  • Re-identification requires minimal effort (database lookup)
  • The data was processed for purposes relating to specific customers

Case 2: Behavioural Profiles Without Direct Identifiers

A profile built from browsing behaviour, purchase patterns, and location data — even without a name or email — constitutes personal data when the profile enables singling out the individual. The Article 29 Working Party in Opinion 4/2007 on the concept of personal data confirmed that "a profile can in itself be sufficient to identify a specific user."

Case 3: Encrypted Data

Encrypted personal data remains personal data for the controller who holds the decryption key. For a third party without the key and no reasonable means to obtain it, the encrypted data may not constitute personal data (applying the Breyer relative approach). However, this assessment must account for future cryptanalytic capabilities.

WP29 Opinion 4/2007 — Key Principles

The Article 29 Working Party Opinion 4/2007 on the concept of personal data established foundational interpretive guidance:

  1. Content element: Information "about" a person exists when the content concerns that individual, regardless of purpose or result
  2. Purpose element: Data is "about" a person when it is used or likely to be used to evaluate, treat, or influence that person
  3. Result element: Data is "about" a person when processing is likely to have an impact on that person's rights or interests
  4. Any one element suffices: Data need only satisfy the content, purpose, OR result element to "relate to" a person

Implementation Procedure for Vanguard Financial Services

Step 1: Data Element Inventory

For each system, catalogue every data element collected, stored, or processed. Record:

  • Field name and data type
  • Source of data (collected from data subject, derived, inferred, received from third party)
  • Sample values (redacted as needed)
  • Current classification if any

Step 2: Apply the Four-Element Test

For each data element, document the assessment against Art. 4(1):

  • Is this any information? (Almost always yes)
  • Does it relate to a natural person? (Content, purpose, or result link)
  • Is the person identified or identifiable? (Direct or indirect, applying Breyer)
  • Is the person a living natural person?

Step 3: Borderline Assessment

For elements not clearly personal or non-personal:

  • Apply the Recital 26 reasonably likely test
  • Consider the Breyer third-party knowledge doctrine
  • Document the reasoning and conclusion
  • Assign to a review schedule (reassess annually or when technology/data partnerships change)

Step 4: Classification Tagging

Apply classification labels:

  • PERSONAL_DIRECT: Directly identifies a natural person
  • PERSONAL_INDIRECT: Indirectly identifies through combination or third-party data
  • SPECIAL_CATEGORY: Art. 9 special category personal data
  • PSEUDONYMISED: Personal data with re-identification key separated
  • ANONYMISED: Verified anonymous data (not personal data)
  • NON_PERSONAL: Not personal data under any reasonable assessment
  • BORDERLINE_REVIEW: Requires periodic reassessment

Step 5: Documentation and Governance

  • Record all classification decisions in the data inventory
  • Link each personal data element to its Art. 6 lawful basis
  • Link Art. 9 special category data to its Art. 9(2) processing condition
  • Schedule annual review of borderline classifications
  • Trigger reclassification when new data partnerships, technologies, or regulatory guidance emerge

Enforcement Precedents

  • Breyer v Bundesrepublik Deutschland (CJEU C-582/14, 2016): Dynamic IP addresses are personal data when legal means to identify exist — established the relative identifiability standard
  • Nowak v Data Protection Commissioner (CJEU C-434/16, 2017): Examination answers and examiner's corrections are personal data of the candidate — applied the broad "relating to" test
  • YS v Minister voor Immigratie (CJEU C-141/12, 2014): Legal analysis in an immigration decision document is personal data of the applicant — the "result" element of the relating-to test
  • Scarlet Extended SA v SABAM (CJEU C-70/10, 2011): IP addresses collected in the context of monitoring internet traffic constitute personal data

Integration Points

  • Art. 9 Special Categories: Personal data classified as special category requires additional processing conditions — see special-category-data skill
  • Art. 30 Records of Processing: Classification feeds directly into the categories of personal data field in RoPA
  • Art. 35 DPIA: High-risk personal data classifications trigger DPIA requirements
  • Art. 25 Data Protection by Design: Classification determines the level of technical protection required

PII Detection in Unstructured Data

Overview

Unstructured data — emails, documents, images, chat logs, call transcripts, and system logs — accounts for an estimated 80% of enterprise data and presents the greatest challenge for privacy compliance. Unlike structured databases where personal data resides in known columns, unstructured data contains PII embedded in free text, attached files, scanned images, and metadata. This skill covers detection approaches using Named Entity Recognition (NER), pattern matching, OCR, and hybrid pipelines, with focus on Microsoft Presidio and spaCy as implementation frameworks.

Unstructured Data Sources at Vanguard Financial Services

SourceVolumePII RiskDetection Challenge
Email (Exchange Online)2.1M messages/monthHIGH — names, account numbers, financial data in body and attachmentsMixed text and attachments; forwarded chains contain accumulated PII
SharePoint documents4.2TB across 1,200 sitesHIGH — contracts, KYC docs, customer correspondenceMultiple formats (docx, pdf, xlsx); embedded images
Teams chat890K messages/monthMEDIUM — casual references to customers, internal discussionsShort messages, abbreviations, context-dependent PII
Application logs50GB/dayMEDIUM — IP addresses, user IDs, error messages with PIIHigh volume, mixed with non-PII technical data
Scanned documents45K pages/monthHIGH — passport scans, signed contracts, medical certificatesRequires OCR; variable image quality
Call transcripts8K transcripts/monthHIGH — customers state names, account numbers, personal detailsSpeech-to-text errors, colloquial language
PDF reports12K documents/monthMEDIUM — financial reports may contain customer listsEmbedded tables, charts with PII labels

Detection Architecture

Microsoft Presidio Pipeline

Presidio is an open-source PII detection and anonymisation SDK developed by Microsoft, designed for integration with enterprise data pipelines.

Input Text/Document
       │
       ▼
┌──────────────────┐
│  Pre-processing  │  Format conversion, encoding normalisation,
│  (text extract)  │  OCR for images/scanned PDFs
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  Presidio        │  Multiple recognisers run in parallel:
│  Analyzer        │  - NER model (spaCy/transformers)
│                  │  - Pattern recognisers (regex)
│                  │  - Custom recognisers (org-specific)
│                  │  - Context-aware enhancers
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  Confidence      │  Each detection assigned confidence score
│  Scoring &       │  Threshold filtering applied
│  Filtering       │  Context enhancement boosts/reduces scores
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  Results         │  PII locations, types, confidence scores
│  (structured)    │  Ready for classification, redaction, or alerting
└──────────────────┘

Component Details

NER Model (spaCy/Transformers):

  • spaCy en_core_web_trf model (transformer-based) for English NER
  • Detects: PERSON, ORG, GPE, DATE, MONEY, CARDINAL entities
  • Custom-trained NER for financial domain entities (IBAN, account numbers)

Pattern Recognisers (Regex):

  • UK National Insurance Number: [A-CEGHJ-PR-TW-Z]{2}\d{6}[A-D]
  • Email address: [a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}
  • UK phone number: (?:0|\+44)\d{10,11}
  • IBAN: [A-Z]{2}\d{2}[A-Z0-9]{4}\d{7}([A-Z0-9]?){0,16}
  • Credit card: \b(?:\d{4}[-\s]?){3}\d{4}\b (with Luhn validation)
  • IP address (v4): \b(?:\d{1,3}\.){3}\d{1,3}\b
  • Date of birth patterns: \b\d{2}[/-]\d{2}[/-]\d{4}\b
  • Vanguard account: VFS-\d{10}
  • ICD-10 code: [A-Z]\d{2}\.\d{1,4}

Context-Aware Enhancement:

  • Keyword proximity: if "National Insurance" appears within 50 characters of a NINO pattern, boost confidence by 15%
  • Section headers: if detected within a section titled "Personal Details" or "Patient Information", boost confidence by 10%
  • Negative context: if pattern appears within code blocks, SQL queries, or configuration files, reduce confidence by 25%

OCR Integration for Scanned Documents

Pipeline for Image-Based PII Detection

Scanned Document / Image
       │
       ▼
┌──────────────────┐
│  Pre-processing  │  Deskew, denoise, contrast enhancement,
│  (image)         │  resolution upscaling (if < 300 DPI)
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  OCR Engine      │  Tesseract OCR (open-source) or
│                  │  Azure AI Document Intelligence (cloud)
│                  │  Output: extracted text with bounding boxes
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  Presidio        │  Standard NER + pattern detection
│  Analyzer        │  on OCR-extracted text
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│  Confidence      │  Adjust for OCR quality:
│  Adjustment      │  OCR confidence < 80% → reduce PII confidence by 20%
│                  │  OCR confidence > 95% → no adjustment
└──────────────────┘

Document Type Detection for Vanguard

Document TypeOCR StrategyExpected PII
Passport scanAzure AI Document Intelligence (ID document model)Full name, DOB, nationality, passport number, photo (biometric)
Utility billGeneral OCR + address pattern recognitionFull name, address, account number
Medical certificateGeneral OCR + health NER modelName, diagnosis, doctor name, dates
Signed contractGeneral OCR + contract template matchingNames, addresses, financial terms, signatures
Cheque imageBanking-specific OCR modelName, account number, sort code, amount

Confidence Scoring Framework

Score Composition

ComponentWeightDescription
Pattern match confidence40%Regex pattern specificity and validation (e.g., Luhn check for credit cards)
NER model confidence30%Model probability score for entity classification
Context enhancement20%Keyword proximity, section header, document type
Source quality10%OCR quality score, document resolution, text extraction confidence

Confidence Thresholds

Confidence LevelScore RangeAction
HIGH85-100%Auto-classify and auto-label; include in discovery report
MEDIUM70-84%Queue for human review; include in discovery report as pending
LOW50-69%Log for audit; do not auto-classify; available for bulk review
BELOW THRESHOLD< 50%Suppress; do not report unless specifically queried

Implementation Patterns

Pattern 1: Email Scanning Pipeline

# Conceptual pipeline for Exchange Online email scanning
# 1. Microsoft Graph API retrieves email messages
# 2. Extract body text (HTML → plain text conversion)
# 3. Extract attachment text (document parsing)
# 4. Run Presidio analyzer on combined text
# 5. Map findings to email metadata (sender, recipients, date)
# 6. Apply classification labels via Microsoft Purview

Pattern 2: Log File Scanning

For application logs, specific patterns dominate:

  • IP addresses in access logs
  • User IDs and session tokens in authentication logs
  • Email addresses in error messages
  • Stack traces containing file paths with usernames

Log scanning requires higher false-positive tolerance and volume-optimised processing.

Pattern 3: Chat Message Scanning

Teams/Slack messages present unique challenges:

  • Short messages with minimal context
  • Abbreviations and informal language
  • Customer names mentioned without surrounding keywords
  • Account numbers shared in fragments across messages

Strategy: scan message threads rather than individual messages to capture context.

Accuracy Benchmarks

Expected Performance by Data Source

SourcePrecision TargetRecall TargetKey Challenges
Email body text> 92%> 88%Forwarded chains, signatures, disclaimers
SharePoint documents (Office formats)> 90%> 85%Embedded tables, headers/footers
Scanned documents (OCR)> 85%> 80%OCR errors, handwriting, poor image quality
Application logs> 88%> 82%IP address over-detection, reference number ambiguity
Chat messages> 80%> 75%Short context, informal language, abbreviations
Call transcripts> 82%> 78%Speech-to-text errors, overlapping speech, accents

Integration Points

  • auto-data-discovery: Unstructured PII detection complements structured data discovery platforms
  • data-labeling-system: Detected PII drives automatic sensitivity label application
  • classification-policy: Detection results feed into classification tier assignment
  • data-inventory-mapping: Unstructured PII findings expand the data inventory to include document repositories and email systems

Pseudonymised vs Anonymised Data Classification

Overview

The distinction between pseudonymised and anonymised data is one of the most consequential classifications in data protection law. Pseudonymised data remains personal data subject to the full GDPR (Art. 4(5), Recital 26). Anonymised data falls outside the GDPR entirely (Recital 26). The boundary between them determines whether processing requires a lawful basis, whether data subjects can exercise rights, and whether the data can be freely shared. This skill provides the analytical framework for making this determination, drawing on the CJEU Breyer ruling (C-582/14), the Article 29 Working Party Opinion 05/2014 on Anonymisation Techniques (WP216), and the ICO motivated intruder test.

Legal Definitions

Pseudonymisation — Art. 4(5)

"The processing of personal data in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately and is subject to technical and organisational measures to ensure that the personal data are not attributed to an identified or identifiable natural person."

Key characteristics:

  • Data has undergone a transformation (tokenisation, hashing, encryption)
  • A re-identification key or method exists
  • The key is kept separately from the data
  • Technical and organisational controls protect the key
  • The data REMAINS personal data under GDPR

Anonymisation — Recital 26

"The principles of data protection should therefore not apply to anonymous information, namely information which does not relate to an identified or identifiable natural person or to personal data rendered anonymous in such a manner that the data subject is not or no longer identifiable."

Key characteristics:

  • No means exist that are reasonably likely to be used for identification
  • Assessment considers ALL means reasonably likely, including third-party means
  • Assessment accounts for cost, time, and available technology
  • Assessment considers technological developments over the data retention period
  • Truly anonymised data is NOT personal data — GDPR does not apply

The Three-Criteria Test — WP29 Opinion 05/2014 (WP216)

The Article 29 Working Party established three criteria for evaluating whether anonymisation is effective. Data is NOT anonymised if ANY of the three attacks can succeed:

Criterion 1: Singling Out

Definition: The possibility of isolating some or all records that identify an individual in the dataset.

Test: Can a specific individual be distinguished from all other individuals in the dataset?

Examples of failure:

  • A dataset with unique combinations of age, postcode, and gender that correspond to only one person in the population
  • A dataset with unique transaction patterns that correspond to a single customer
  • Removing a name but retaining a rare medical condition in a small geographic area

Mitigation techniques: k-anonymity (ensure at least k individuals share each combination of quasi-identifiers), aggregation to sufficiently large groups

Criterion 2: Linkability

Definition: The ability to link at least two records concerning the same data subject or a group of data subjects, either within the same dataset or between two separate datasets.

Test: Can records be combined across datasets to build a profile of an individual?

Examples of failure:

  • A hashed email address that can be matched across two datasets released by the same controller
  • A dataset with date of birth, postcode, and hospital visit date that can be linked to public voter registration records
  • Transaction patterns that can be matched with social media activity timestamps

Mitigation techniques: l-diversity, t-closeness, differential privacy noise addition, removing linkable quasi-identifiers

Criterion 3: Inference

Definition: The possibility of deducing, with significant probability, the value of an attribute from the values of a set of other attributes.

Test: Can sensitive information about an individual be inferred from the released data?

Examples of failure:

  • A dataset shows that all patients in a particular age-postcode group have the same diagnosis
  • Salary data for a small department where only one person holds a particular role
  • Aggregate statistics with small cell sizes that reveal individual values

Mitigation techniques: t-closeness (ensure attribute value distribution in each equivalence class is close to the overall distribution), differential privacy, cell suppression for small groups

Assessment Matrix

CriterionPasses?Overall Anonymisation Status
Singling out: NOT possibleLinkability: NOT possibleInference: NOT possible → ANONYMISED
Singling out: possibleAnyAny → NOT anonymised (personal data)
AnyLinkability: possibleAny → NOT anonymised (personal data)
AnyAnyInference: possible → NOT anonymised (personal data)

All three criteria must be satisfied for data to qualify as anonymised.

The Motivated Intruder Test (ICO)

The ICO supplements the WP29 framework with the "motivated intruder" test:

The motivated intruder is:

  • A person who starts without any prior knowledge of the data subject
  • Has access to the dataset and to publicly available resources (internet, public registers, social media, published statistics)
  • Has reasonable competence and motivation to re-identify (not a specialist)
  • Does NOT have access to specialist equipment, techniques, or insider knowledge
  • Would invest reasonable time and effort (days, not months)

If the motivated intruder could re-identify any individual: the data is NOT anonymised.

This test provides a practical complement to the WP29 criteria by establishing a baseline attacker model for the "reasonably likely means" assessment under Recital 26.

Anonymisation Techniques — Effectiveness Assessment

Technique 1: Randomisation (Noise Addition)

ApproachDescriptionSingling OutLinkabilityInference
Noise additionAdd random values to numerical attributesReducedReducedReduced
PermutationShuffle values across records within groupsMitigated (within group)ReducedReduced
Differential privacyAdd calibrated noise to query resultsProvably mitigated (with epsilon budget)Provably mitigatedProvably mitigated

Technique 2: Generalisation

ApproachDescriptionSingling OutLinkabilityInference
k-anonymityEnsure each combination of quasi-identifiers applies to at least k recordsMitigated (k ≥ 5 recommended)Not addressedNot addressed
l-diversityWithin each k-anonymous group, ensure at least l distinct values for sensitive attributesMitigatedPartially mitigatedReduced
t-closenessEnsure distribution of sensitive attribute in each group is within distance t of overall distributionMitigatedReducedMitigated
AggregationReplace individual records with group statisticsMitigated (if groups large enough)MitigatedResidual risk with small groups

Technique 3: Suppression

ApproachDescriptionSingling OutLinkabilityInference
Record suppressionRemove outlier recordsReduces unique recordsNo effect on remainingNo effect on remaining
Cell suppressionReplace small cell values with null/rangeReduces singling out for small groupsNo direct effectReduces inference from small cells
Attribute suppressionRemove entire columns (direct identifiers, high-risk quasi-identifiers)Reduces identifiabilityReduces cross-dataset linkageNo direct effect

WP29 Effectiveness Summary (from Opinion 05/2014, Section 4)

TechniqueSingling OutLinkabilityInferenceRecommended Use
Pseudonymisation (tokenisation, hashing)Still possibleStill possibleStill possibleNOT anonymisation — remains personal data
k-anonymity alonePartially mitigatedNOT mitigatedNOT mitigatedInsufficient alone
l-diversityMitigatedPartially mitigatedPartially mitigatedStronger than k-anonymity
Differential privacyProvably mitigatedProvably mitigatedProvably mitigatedStrongest formal guarantee
Aggregation + suppressionMitigatedMitigatedResidualSuitable for statistical outputs

Decision Framework for Classification

Data has undergone a de-identification process
  │
  ├─► Step 1: Does a re-identification key exist?
  │     YES → PSEUDONYMISED (personal data, GDPR applies)
  │     NO  → Continue to Step 2
  │
  ├─► Step 2: Apply WP29 Three-Criteria Test
  │     2a. Singling out test — can any individual be isolated?
  │     2b. Linkability test — can records be linked to an individual?
  │     2c. Inference test — can attributes be inferred for an individual?
  │
  │     ANY test fails → NOT anonymised (personal data, GDPR applies)
  │     ALL tests pass → Continue to Step 3
  │
  ├─► Step 3: Apply Motivated Intruder Test (ICO)
  │     Could a motivated person with public resources re-identify anyone?
  │     YES → NOT anonymised
  │     NO  → Continue to Step 4
  │
  ├─► Step 4: Breyer Relative Assessment
  │     Does the specific data controller have legal means to obtain
  │     complementary data enabling identification? (per C-582/14)
  │     YES → NOT anonymised FOR THIS CONTROLLER
  │     NO  → Continue to Step 5
  │
  └─► Step 5: Technology Projection
        Over the intended retention period, are technological developments
        reasonably anticipated that would enable re-identification?
        YES → NOT anonymised (or must be re-assessed before technology matures)
        NO  → ANONYMISED — GDPR does not apply

        CLASSIFICATION: ANONYMISED
        Document: technique used, assessment reasoning, review date,
        assessor, technology projection horizon

CJEU Case C-413/23 — European Data Protection Supervisor v Single Resolution Board (Pending)

In 2023, the CJEU was asked to rule on whether pseudonymised data transmitted to a recipient who does not have the re-identification key constitutes personal data for that recipient. Advocate General Szpunar's opinion (September 2024) suggested a relative approach — if the recipient cannot reasonably re-identify, the data may not be personal data for them. This case, when decided, may significantly impact the pseudonymisation vs anonymisation boundary by establishing that the same dataset can be pseudonymised (personal data) for the data holder and effectively anonymous for a recipient without the key.

Implementation at Vanguard Financial Services

Common Pseudonymisation Scenarios

ScenarioTechniqueClassificationReasoning
Customer analytics dataset with account IDs replaced by random tokensTokenisationPSEUDONYMISEDVanguard holds the token-to-account mapping table
Employee survey responses with employee IDs removed but department and role retainedAttribute suppression + quasi-identifiers remainPSEUDONYMISED (likely)Unique department-role combinations may enable singling out for small departments
Aggregated website traffic statistics: page views by week, no user identifiersAggregationANONYMISED (if cell sizes adequate)No individual can be singled out from aggregate page view counts with >1000 visitors per cell
Hashed customer emails for analyticsHashing (SHA-256)PSEUDONYMISEDHash can be reversed via dictionary attack or rainbow table; same email always produces same hash, enabling linkage

Anonymisation Validation Process

Before classifying data as anonymised, Vanguard requires:

  1. Written anonymisation assessment documenting technique applied and WP29 three-criteria evaluation
  2. Motivated intruder test conducted by Privacy Engineering team
  3. DPO review and sign-off
  4. Annual reassessment (technology evolution check)
  5. If data is shared externally: assessment from recipient's perspective (Breyer relative approach)

Enforcement Precedents

  • Breyer v Bundesrepublik Deutschland (CJEU C-582/14, 2016): Established relative approach to identifiability — central to distinguishing pseudonymised (identifiable) from anonymised (not identifiable) for a specific controller
  • CNIL — Practical Guide on Anonymisation (2019): CNIL rejected k-anonymity alone as sufficient for anonymisation, requiring assessment against all three WP29 criteria
  • Austrian DPA — 1&1 Telecom case (2019): EUR 9.55 million fine partly related to inadequate pseudonymisation measures — telephone customer service could identify callers with minimal verification, demonstrating pseudonymisation was ineffective

Integration Points

  • personal-data-test: Pseudonymised data is personal data; anonymised data is not — classification feeds the Art. 4(1) test
  • classification-policy: Pseudonymised data classified as Internal or Confidential; anonymised data classified as Public or Internal
  • ai-training-data-class: AI training datasets frequently use pseudonymisation; this skill validates whether the technique achieves true anonymisation
  • data-lineage-tracking: Lineage must track the anonymisation transformation and preserve the assessment record

Special Category Data Classification — GDPR Art. 9

Overview

Article 9(1) of the GDPR establishes a general prohibition on processing special categories of personal data. These categories were identified by the European legislature as carrying heightened risk to fundamental rights and freedoms due to their potential for discrimination, social stigma, or irreversible harm. Processing is permitted only when one of the ten conditions in Art. 9(2)(a)-(j) is satisfied, in addition to a valid lawful basis under Art. 6. This skill provides a systematic framework for identifying special category data across enterprise systems and mapping each instance to an appropriate processing condition.

The Eight Special Categories — Art. 9(1)

Category 1: Racial or Ethnic Origin

Definition: Data revealing or from which racial or ethnic origin can be inferred, including direct declarations, photographs, names characteristic of particular ethnic groups, or nationality data when used as a proxy for ethnic origin.

Examples at Vanguard Financial Services:

  • Employee self-identification forms for diversity monitoring
  • Customer photographs in KYC documentation
  • Nationality fields when used to infer ethnicity
  • Language preference data when combined with geographic data to infer origin

Boundary Cases:

  • Nationality alone is not necessarily racial/ethnic origin data, but becomes special category when processed for diversity analysis (EDPB, WP251rev.01)
  • Photographs are special category only when processed through facial recognition for the purpose of uniquely identifying — a passport photo stored for KYC is not biometric data, but the same photo run through facial comparison software becomes biometric data

Category 2: Political Opinions

Definition: Data revealing political views, party membership, voting behaviour, political donations, or participation in political activities.

Examples at Vanguard Financial Services:

  • Employee political donation records (where legally collected for compliance with US PAC reporting requirements)
  • Customer complaints referencing political views
  • Social media data processed for marketing that reveals political affiliations

Key Precedent: Austrian Post (Österreichische Post AG) — Austrian DPA and subsequently CJEU Case C-300/21 (2023) — fined EUR 18 million for processing political affinity scores derived from statistical models applied to demographic data. The CJEU confirmed that data revealing political opinions includes inferred data, not only data directly provided by the data subject.

Category 3: Religious or Philosophical Beliefs

Definition: Data revealing religious faith, atheism, agnosticism, philosophical convictions, or related practices. Includes dietary preferences when they indicate religious observance (halal, kosher), religious holiday requests, and membership of religious organisations.

Examples at Vanguard Financial Services:

  • Employee religious holiday accommodation requests
  • Dietary restriction data for corporate event catering
  • Charitable donation patterns to religious organisations (from transaction data)

Category 4: Trade Union Membership

Definition: Data revealing whether an individual is or was a member of a trade union. Includes union dues deductions from payroll, attendance at union meetings, and communications with union representatives.

Examples at Vanguard Financial Services:

  • Payroll deduction records for union dues
  • HR records of union membership status
  • Meeting room booking records for union activities

Category 5: Genetic Data — Art. 4(13)

Definition: Personal data relating to the inherited or acquired genetic characteristics of a natural person which give unique information about the physiology or health of that natural person, resulting in particular from an analysis of a biological sample from the natural person in question.

Examples at Vanguard Financial Services:

  • Genetic test results in employee wellness programs (if offered)
  • Insurance underwriting data derived from genetic testing (prohibited in many jurisdictions)
  • Biobank research data from corporate health initiatives

Regulatory Note: The Genetic Information Nondiscrimination Act (GINA) in the US prohibits use of genetic information in health insurance and employment. In the EU, genetic data receives dual protection under both Art. 9 and specific Member State genetic data legislation.

Category 6: Biometric Data — Art. 4(14)

Definition: Personal data resulting from specific technical processing relating to the physical, physiological, or behavioural characteristics of a natural person, which allow or confirm the unique identification of that natural person, such as facial images or dactyloscopic data.

Critical Distinction: Biometric data is special category ONLY when processed "for the purpose of uniquely identifying a natural person" (Art. 9(1)). A photograph stored in an HR file is personal data but not special category. The same photograph processed through facial recognition software for access control is special category biometric data.

Examples at Vanguard Financial Services:

  • Fingerprint scanners for building access control
  • Facial recognition for secure facility entry
  • Voice recognition for telephone banking authentication
  • Behavioural biometrics (keystroke dynamics) for fraud detection

Key Precedent: Clearview AI — CNIL Decision SAN-2022-019 (20 October 2022) — EUR 20 million fine for processing biometric data (facial recognition) without lawful basis and without conducting DPIA.

Category 7: Health Data — Recital 35

Definition: Personal data related to the physical or mental health of a natural person, including the provision of health care services, which reveal information about their health status. Recital 35 specifies this includes data pertaining to the health status of a data subject which reveals information relating to the past, current, or future physical or mental health of the data subject, including: registration for health care services, number/symbol/identifier assigned for health purposes, information derived from testing or examination of a body part or bodily substance, and any information on a disease, disability, disease risk, medical history, clinical treatment, or physiological or biomedical condition.

Examples at Vanguard Financial Services:

  • Employee sick leave records and medical certificates
  • Disability accommodation requests
  • Occupational health assessment results
  • Health insurance claims data
  • Wellness program participation data (step counts, health metrics)
  • COVID-19 vaccination status records
  • Drug and alcohol testing results
  • Ergonomic assessment data indicating physical conditions

Breadth of Health Data: The CJEU has interpreted health data broadly. In Case C-184/20 (Vyriausioji tarnybinės etikos komisija, 2022), the Court held that data which indirectly reveals health information (such as a spouse's name in a declaration of interests that could reveal sexual orientation or health status) may constitute special category data.

Category 8: Sex Life or Sexual Orientation

Definition: Data concerning sexual behaviour, sexual preferences, or sexual orientation. Includes data from which sexual orientation can be inferred.

Examples at Vanguard Financial Services:

  • Employee diversity monitoring data including sexual orientation
  • Beneficiary designations that reveal same-sex partnerships
  • Customer data from which sexual orientation can be inferred (partner names, household composition)

Processing Conditions — Art. 9(2)(a)-(j)

Processing of special category data is lawful ONLY when one of these conditions is met (in addition to Art. 6 lawful basis):

ConditionArt. 9(2)RequirementsVanguard Application
Explicit consent(a)Must be freely given, specific, informed, unambiguous, and EXPLICIT (higher standard than Art. 6(1)(a) consent). Must be a clear affirmative statement, not implied.Employee diversity monitoring with opt-in explicit consent
Employment and social security law(b)Processing necessary for obligations under employment, social security, or social protection law. Must be authorised by EU or Member State law or collective agreement with appropriate safeguards.Payroll processing of trade union dues, occupational health assessments required by law
Vital interests(c)Processing necessary to protect vital interests where data subject is physically or legally incapable of giving consent.Emergency medical situations where employee is incapacitated
Legitimate activities of non-profit(d)Processing by foundation, association, or not-for-profit body with political, philosophical, religious, or trade union aims, relating to members or regular contacts, with no disclosure outside the body without consent.Not applicable to Vanguard (commercial entity)
Data manifestly made public(e)Data subject has manifestly made the data public (e.g., publicly declared political views on social media, public disclosure of health condition).Customer data voluntarily posted on public forums
Legal claims(f)Processing necessary for establishment, exercise, or defence of legal claims, or whenever courts are acting in their judicial capacity.Litigation holds involving health data in employment disputes
Substantial public interest(g)Processing necessary for reasons of substantial public interest, on basis of EU or Member State law, proportionate to the aim, with appropriate safeguards.Regulatory reporting obligations (e.g., AML suspicious activity involving special category data)
Health care and occupational medicine(h)Processing necessary for preventive or occupational medicine, assessment of working capacity, medical diagnosis, health/social care provision, or management of health systems. Must be processed by or under responsibility of a professional with secrecy obligation.Occupational health surveillance mandated by workplace health regulations
Public health(i)Processing necessary for public health reasons such as protection against serious cross-border threats to health, ensuring high standards of quality and safety for medicines/medical devices.COVID-19 workplace safety measures (now largely wound down)
Archiving, research, statistics(j)Processing necessary for archiving in the public interest, scientific or historical research, or statistical purposes under Art. 89(1), with appropriate safeguards including data minimisation.Internal workforce diversity statistical analysis

Identification Methodology

Automated Detection Indicators

Indicator TypeDetection ApproachSpecial Category Flag
Field names containing health terminologyRegex pattern matching: `diagnosissymptom
ICD-10/ICD-11 codesCode format detection: [A-Z][0-9]{2}(\.[0-9]{1,4})?Health data
Biometric template formatsBinary header detection for ISO 19795, ANSI/INCITS 378 fingerprint templatesBiometric data
Diversity form fieldsField labels matching: `ethnicityrace
Genetic marker identifiersSNP identifiers (rs-numbers), gene names (HUGO nomenclature)Genetic data

Manual Review Triggers

Processing activities that warrant manual special category review:

  1. Any processing involving employee HR data (may contain health, union, diversity data)
  2. Customer onboarding with identity verification (photographs may become biometric)
  3. Marketing analytics using behavioural profiling (may infer political views, health status)
  4. Insurance or benefits processing (health data inherent)
  5. AI/ML training using customer or employee data (risk of inferring special categories)

Data Protection Impact Assessment Requirement

Under Art. 35(3)(b), processing special category data on a large scale automatically triggers a mandatory DPIA. For Vanguard Financial Services:

Processing ActivityScale AssessmentDPIA Required?
Employee health records12,000 employees — large scale for employerYES
Fingerprint access controlAll office buildings, 8,500 usersYES
Customer KYC photographs2.4 million customersYES (if facial recognition applied)
Diversity monitoring surveyVoluntary, ~3,000 respondentsYES (special category + employment context)

Enforcement Precedents

  • Österreichische Post AG (Austrian DPA / CJEU C-300/21, 2023): EUR 18 million fine for processing inferred political affinity scores. Established that data need not be directly provided by the data subject to constitute special category data — statistical inference suffices.
  • Clearview AI (CNIL SAN-2022-019, 2022): EUR 20 million fine for biometric data processing without lawful basis or DPIA.
  • Asociación Profesional Elite Taxi v Uber Systems Spain (CJEU C-434/15, 2017): While primarily an internal market case, the opinion addressed processing of driver data that could reveal trade union membership.
  • Hungarian NAIH (2021): Fined a retailer HUF 100 million for processing employee health data (COVID-19 test results) beyond the scope authorised by public health legislation.
  • Swedish DPA (2019): Fined a school SEK 200,000 for using facial recognition (biometric data) for student attendance tracking without valid legal basis under Art. 9(2).

Integration Points

  • personal-data-test: Special category classification presupposes personal data classification — Art. 9 applies only to data that is first personal data under Art. 4(1)
  • criminal-data-handling: Art. 10 criminal data has separate rules but is often processed alongside special categories
  • classification-policy: Special category designation drives the highest classification tier (Restricted)
  • data-inventory-mapping: Art. 30 RoPA must identify special category data per processing activity

Sponsors

Details

Marketplace
mukul975/Privacy-Data-Protection-Skills
Source
plugins/data-classification-skills
Author
Mahipal
Category
Engineering
License
Apache-2.0
Keywords
data-classification, pii, data-inventory, data-lineage

More from mukul975/Privacy-Data-Protection-Skills

All 21