
privacy-by-design-skills
13 privacy-by-design skills: data minimization architecture, federated learning, homomorphic encryption, SMPC, PETs, LINDDUN
- version
- 1.0.0
- GitHub stars
- 340
Install
/plugin install privacy-by-design-skills@privacy-data-protection-skillsRun it inside Claude Code after adding the marketplace with /plugin marketplace add mukul975/Privacy-Data-Protection-Skills.
Contents
Discovered from sourceSkills 13
Applying Privacy Design Patterns
Overview
Privacy design patterns provide reusable architectural solutions for implementing data protection principles in system design. Jaap-Henk Hoepman's framework (2014, expanded in "Privacy Design Strategies: The Eight Strategies for GDPR Compliance") defines eight privacy design strategies organized into two categories: data-oriented strategies (minimize, hide, separate, abstract) that focus on the processing of personal data itself, and process-oriented strategies (inform, control, enforce, demonstrate) that focus on the organizational processes surrounding data processing.
These patterns directly implement GDPR Article 25(1) data protection by design and map to specific GDPR principles under Article 5.
The Eight Privacy Design Patterns
Data-Oriented Strategies
1. MINIMIZE
Principle: Limit the processing of personal data as much as possible.
GDPR mapping: Article 5(1)(c) data minimization, Article 25(2) by default.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Select before collect | Determine which data is necessary before designing collection interfaces | API allowlists, form field audits |
| Exclude | Remove unnecessary data elements from collection | Schema validation rejecting non-required fields |
| Strip | Remove identifying information as soon as possible after collection | Pseudonymization at ingestion boundary |
| Destroy | Delete data as soon as the purpose is fulfilled | TTL-based automated deletion |
Prism Data Systems AG Implementation:
The customer onboarding API at Prism Data Systems AG validates incoming requests against a strict allowlist. The /api/v2/register endpoint accepts only email, display_name, and country_code. The date_of_birth field is collected only during age verification and is converted to a boolean is_age_verified within 24 hours, with the raw date destroyed.
2. HIDE
Principle: Protect personal data, or make it unlinkable or unobservable.
GDPR mapping: Article 5(1)(f) integrity and confidentiality, Article 32(1)(a) encryption and pseudonymisation.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Encrypt | Apply cryptographic protection to data at rest and in transit | AES-256-GCM field-level encryption, TLS 1.3 |
| Hash | Replace identifiers with irreversible digests | SHA-256 for log anonymization |
| Mix | Combine data from multiple subjects to prevent singling out | k-anonymity, differential privacy noise |
| Obfuscate | Add noise or perturbation to prevent precise inference | Differential privacy, data masking |
| Dissociate | Break the link between data and identity | Pseudonymization with separated key storage |
Prism Data Systems AG Implementation: All personally identifiable fields are encrypted with per-field AES-256-GCM Data Encryption Keys (DEKs) managed in AWS KMS. Customer identifiers entering the analytics pipeline are pseudonymized via HMAC-SHA256 with keys stored in a Hardware Security Module physically separated from analytics infrastructure.
3. SEPARATE
Principle: Process personal data in a distributed fashion, preventing correlation.
GDPR mapping: Article 5(1)(b) purpose limitation, Article 25(1) by design.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Isolate | Process different categories of data in separate systems | Purpose-partitioned databases, microservice isolation |
| Distribute | Spread data across multiple locations to prevent single-point access | Federated learning, SMPC, sharding |
Prism Data Systems AG Implementation: Prism Data Systems AG maintains purpose-partitioned PostgreSQL databases: the authentication database stores only credential-related data, the billing database stores only financial data, and the analytics warehouse stores only pseudonymized event data. No single database contains a complete profile of any customer. Cross-purpose joins require explicit compatibility assessment per Article 6(4).
4. ABSTRACT
Principle: Limit the detail of personal data as much as possible.
GDPR mapping: Article 5(1)(c) data minimization, Recital 26 anonymization.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Summarize | Replace detailed data with aggregated summaries | Aggregate reporting with minimum group size 11 |
| Group | Generalize values to broader categories | Age ranges instead of exact ages, region instead of postal code |
| Perturb | Add randomness to exact values | Differential privacy, random rounding |
Prism Data Systems AG Implementation: Customer age is stored as a 5-year bracket (e.g., "25-29") rather than exact date of birth. Geographic data is generalized from full postal code to canton-level. Analytics dashboards enforce a minimum group size of 11 records per cell, with smaller groups suppressed and displayed as "< 11."
Process-Oriented Strategies
5. INFORM
Principle: Inform data subjects about the processing of their personal data.
GDPR mapping: Articles 12-14 transparency obligations.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Supply | Proactively provide privacy information | Layered privacy notices, just-in-time notifications |
| Notify | Alert data subjects about processing events | Email notifications for new data access, breach notifications |
| Explain | Provide meaningful explanations of processing logic | Explainable AI outputs, processing purpose descriptions |
Prism Data Systems AG Implementation: A layered privacy notice is presented at every data collection point. The first layer is a plain-language summary (Flesch-Kincaid grade 8). The second layer provides full Article 13 information. Just-in-time notifications appear when a new feature requires additional data: "To activate bulk export, Prism Data Systems AG needs to process your API usage history."
6. CONTROL
Principle: Provide data subjects with control over the processing of their personal data.
GDPR mapping: Articles 15-22 data subject rights, Article 7(3) consent withdrawal.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Consent | Obtain and manage consent for each processing purpose | Granular consent UI with unticked defaults |
| Choose | Allow data subjects to select processing options | Privacy preference center |
| Update | Enable data subjects to correct their data | Self-service profile editing |
| Retract | Enable data subjects to withdraw consent or request erasure | One-click consent withdrawal, erasure request workflow |
Prism Data Systems AG Implementation:
The Privacy Preference Center at account.prism-data.ch/privacy provides data subjects with granular controls: per-purpose consent toggles, data download (portability), correction interface, and one-click erasure request. Consent withdrawal takes effect within 24 hours and triggers downstream processing cessation.
7. ENFORCE
Principle: Commit to processing personal data in a privacy-friendly way, and enforce this.
GDPR mapping: Article 24 controller responsibility, Article 25(1) by design, Article 28 processor contracts.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Create | Define and publish privacy policies and standards | GDPR policy framework, data classification standard |
| Maintain | Regularly update and enforce privacy policies | Quarterly policy reviews, automated compliance checks |
| Uphold | Implement technical enforcement of privacy rules | Purpose-based access control (OPA), automated retention |
Prism Data Systems AG Implementation: Privacy policies are enforced technically through Open Policy Agent (OPA) rules that evaluate every data access request against the purpose registry. The CI/CD pipeline includes a privacy gate that blocks deployment of services failing the data minimization assessment. Processor contracts include standard contractual clauses reviewed semi-annually.
8. DEMONSTRATE
Principle: Demonstrate compliance with privacy policies and applicable regulations.
GDPR mapping: Article 5(2) accountability, Article 30 records of processing, Article 35 DPIA.
Sub-patterns:
| Sub-pattern | Description | Implementation |
|---|---|---|
| Record | Maintain comprehensive records of processing activities | Article 30 register, consent logs, DPIA repository |
| Audit | Conduct regular privacy audits | Quarterly internal audits, annual external audit |
| Report | Generate compliance reports for regulators and management | DPO quarterly report, board privacy report |
Prism Data Systems AG Implementation: All data access events are logged in an immutable audit trail with: timestamp, requester identity, purpose declaration, data categories accessed, and authorization decision. The Article 30 register is maintained as a living document updated within 5 business days of any processing change. DPIAs are conducted for all high-risk processing and reviewed annually.
Pattern Selection Methodology
Step 1: Map Processing to GDPR Principles
For each processing activity, identify which GDPR principles are most relevant:
| GDPR Principle | Primary Pattern | Supporting Patterns |
|---|---|---|
| Data minimization (Art. 5(1)(c)) | MINIMIZE | ABSTRACT, HIDE |
| Purpose limitation (Art. 5(1)(b)) | SEPARATE | ENFORCE |
| Storage limitation (Art. 5(1)(e)) | MINIMIZE (Destroy) | ENFORCE |
| Integrity & confidentiality (Art. 5(1)(f)) | HIDE | ENFORCE |
| Transparency (Art. 5(1)(a)) | INFORM | DEMONSTRATE |
| Lawfulness (Art. 5(1)(a)) | CONTROL | ENFORCE |
| Accuracy (Art. 5(1)(d)) | CONTROL (Update) | INFORM |
| Accountability (Art. 5(2)) | DEMONSTRATE | ENFORCE |
Step 2: Assess Pattern Applicability
Score each pattern's applicability (1-5) for the specific processing activity.
Step 3: Select and Combine
Select the highest-scoring patterns. Most processing activities require a combination of 3-5 patterns applied at different architectural layers.
Key Regulatory References
- GDPR Article 5 — Principles relating to processing of personal data
- GDPR Article 25(1) — Data protection by design
- GDPR Article 25(2) — Data protection by default
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
- Hoepman, J.-H. (2014). "Privacy Design Strategies." IFIP SEC. Extended in "Privacy Design Strategies: The Eight Strategies for GDPR Compliance" (2022).
Automating Storage Limitation Controls
Overview
Article 5(1)(e) of the GDPR requires that personal data be "kept in a form which permits identification of data subjects for no longer than is necessary for the purposes for which the personal data are processed." This is the storage limitation principle. Controllers must define retention periods for each category of personal data based on the purpose of processing (Article 13(2)(a) requires disclosure of retention periods or criteria).
Automation is critical because manual retention management at scale inevitably leads to compliance gaps. The EDPB Guidelines 4/2019 on Article 25 specifically identify automated deletion mechanisms as an example of data protection by design.
Retention Policy Framework
Retention Period Determination
Each retention period must be justified by reference to the processing purpose, legal requirements, or legitimate business need:
| Data Category | Purpose | Lawful Basis | Retention Period | Retention Trigger | Legal Reference |
|---|---|---|---|---|---|
| Customer account data | Service delivery | Art. 6(1)(b) | Duration of contract + 90 days | Account closure | Art. 5(1)(e), Art. 17(1)(a) |
| Transaction records | Tax compliance | Art. 6(1)(c) | 10 years | Transaction date | Swiss CO Art. 958f, EU VAT Directive 2006/112/EC Art. 242 |
| Support tickets | Customer service | Art. 6(1)(b) | 3 years | Ticket closure | Art. 5(1)(e), limitation period |
| Server access logs | Security monitoring | Art. 6(1)(f) | 12 months | Log creation date | Art. 5(1)(e), ENISA recommendation |
| Marketing consent records | Accountability | Art. 6(1)(c) | Duration of consent + 5 years | Consent withdrawal | Art. 7(1), Art. 5(2) |
| Employment records | Legal obligation | Art. 6(1)(c) | Duration of employment + 10 years | Employment end | Swiss OR Art. 127-128 |
| Cookie analytics | Product improvement | Art. 6(1)(a) | 13 months | Data collection | ePrivacy Directive Art. 5(3), CNIL guidance |
| DPIA records | Accountability | Art. 6(1)(c) | Duration of processing + 3 years | Processing cessation | Art. 35, Art. 5(2) |
TTL-Based Deletion Architecture
┌─────────────────────────────────────────────────────────────┐
│ Retention Policy Engine │
│ ┌───────────────┐ ┌────────────────┐ ┌───────────────┐ │
│ │ Policy Store │ │ TTL Calculator │ │ Legal Hold │ │
│ │ (per category)│ │ (trigger+period)│ │ Registry │ │
│ └───────┬───────┘ └───────┬────────┘ └──────┬────────┘ │
│ └──────────────────┼───────────────────┘ │
└─────────────────────────────┼───────────────────────────────┘
│
┌─────────────────────────────▼───────────────────────────────┐
│ Lifecycle Scheduler │
│ ┌──────────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Expiry Scanner │ │ Archive Mover│ │ Deletion │ │
│ │ (daily cron) │ │ (pre-delete) │ │ Executor │ │
│ └──────────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────┬───────────────────────────────┘
│
┌─────────────────────────────▼───────────────────────────────┐
│ Audit and Verification │
│ ┌──────────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Deletion Log │ │ Verification │ │ Exception │ │
│ │ (immutable) │ │ Scanner │ │ Reporter │ │
│ └──────────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────────┘
Deletion Patterns for Distributed Systems
| Pattern | Description | Use Case | Complexity |
|---|---|---|---|
| TTL columns | Database column storing expiry timestamp; background job deletes expired rows | Single-database systems, Redis, DynamoDB | Low |
| Event-driven deletion | Lifecycle event (account closure) publishes deletion command to message queue | Microservice architectures | Medium |
| Tombstone + compaction | Mark records as deleted; background compaction physically removes data | Append-only stores, event sourcing | Medium |
| Cascade deletion | Delete propagates across foreign key relationships | Relational databases with referential integrity | Medium |
| Distributed saga | Coordinated deletion across multiple services with rollback on failure | Polyglot persistence, cross-service data | High |
| Crypto-shredding | Destroy the encryption key rendering the data unreadable without physical deletion | Systems where physical deletion is impractical (backups, immutable logs) | Medium |
Legal Hold Management
Legal holds suspend automated deletion when data is subject to litigation, regulatory investigation, or audit proceedings.
Legal Hold Lifecycle:
- Hold initiation — Legal counsel issues a hold notice specifying the scope (data categories, date range, affected subjects) and the matter reference
- Hold registration — The Retention Policy Engine registers the hold, creating an exemption that overrides TTL-based deletion for matching records
- Hold monitoring — Monthly review by legal counsel to confirm the hold remains necessary
- Hold release — Legal counsel issues a release notice; the Retention Policy Engine removes the exemption and resumes normal retention processing
- Catch-up deletion — Records that would have been deleted during the hold period are now evaluated against current retention policies and deleted if expired
Prism Data Systems AG Implementation:
| Hold Parameter | Value |
|---|---|
| Hold scope format | JSON document specifying data_categories, date_range_start, date_range_end, subject_identifiers, matter_reference |
| Hold registry | PostgreSQL table with immutable audit trail |
| Review cadence | Monthly by legal counsel, quarterly by DPO |
| Maximum hold duration without renewal | 12 months |
| Hold release approval | Dual approval: legal counsel + DPO |
Archival Policies
Three-Tier Storage Lifecycle
| Tier | Storage Type | Access Pattern | Encryption | Retention Phase |
|---|---|---|---|---|
| Hot | Primary database (PostgreSQL) | Real-time read/write | AES-256-GCM field-level | Active processing |
| Warm | Object storage (S3-compatible, Glacier Instant Retrieval class) | Occasional read, no write | AES-256-GCM object-level | Post-purpose, pre-deletion (legal/audit) |
| Cold | Offline archive (tape/vault) | Restore on request (24-48h) | AES-256-GCM + envelope encryption | Statutory retention only |
Transition Rules for Prism Data Systems AG:
| Transition | Trigger | Automation |
|---|---|---|
| Hot → Warm | Purpose fulfilled (e.g., account closed) + 90 days | Daily lifecycle job |
| Warm → Cold | Warm retention period expires, statutory retention still applies | Monthly lifecycle job |
| Cold → Deletion | Statutory retention period expires, no legal hold active | Quarterly lifecycle job with DPO approval gate |
| Any → Legal Hold | Legal hold registered for matching scope | Immediate suspension |
Implementation Workflow
-
Inventory data stores — Catalog every system that stores personal data, including primary databases, caches, message queues, logs, backups, and third-party integrations.
-
Map retention requirements — For each data category and processing purpose, determine the retention period based on purpose necessity, legal obligations, and limitation periods.
-
Implement TTL metadata — Add retention-related columns or metadata to each data store:
created_at,retention_trigger_event,retention_trigger_date,ttl_expiry_date,legal_hold_id. -
Deploy lifecycle scheduler — Implement a background service that scans for expired records daily, checks for active legal holds, and executes deletion or archival transitions.
-
Implement deletion verification — After each deletion cycle, run a verification scan to confirm that all targeted records have been removed from primary stores, replicas, and caches.
-
Configure backup rotation — Ensure that backup retention aligns with the longest applicable retention period. Implement crypto-shredding for backup media where granular deletion is impractical.
-
Deploy monitoring and alerting — Monitor deletion job success rates, flag records that exceed their TTL by more than 7 days, and alert on legal hold expirations.
Key Regulatory References
- GDPR Article 5(1)(e) — Storage limitation principle
- GDPR Article 13(2)(a) — Retention period disclosure requirement
- GDPR Article 17(1)(a) — Right to erasure when data no longer necessary
- GDPR Article 25(1) — Data protection by design
- GDPR Recital 39 — Storage limitation guidance
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
- Swiss Code of Obligations Art. 958f — 10-year retention for business records
- EU VAT Directive 2006/112/EC Art. 242 — Invoice retention requirements
Building Purpose Limitation Enforcement
Overview
Article 5(1)(b) of the GDPR requires that personal data be "collected for specified, explicit and legitimate purposes and not further processed in a manner that is incompatible with those purposes." This is the purpose limitation principle, one of the foundational data protection principles.
The controller must specify the purpose at or before the time of collection (Article 13(1)(c) for direct collection, Article 14(1)(c) for indirect collection). Any subsequent processing for a different purpose requires either a new lawful basis or must pass the compatibility assessment under Article 6(4).
The Article 29 Working Party Opinion 03/2013 on purpose limitation (WP203) provides detailed guidance on assessing compatibility, identifying five key factors codified in Article 6(4).
Article 6(4) Compatibility Assessment Factors
When a controller wishes to process personal data for a purpose other than that for which it was collected, Article 6(4) requires an assessment of compatibility considering:
| Factor | Article 6(4) Reference | Assessment Question |
|---|---|---|
| Link between purposes | Article 6(4)(a) | Is there a connection between the original and new purpose? |
| Context of collection | Article 6(4)(b) | What is the relationship between controller and data subject? What are the reasonable expectations? |
| Nature of data | Article 6(4)(c) | Does the data include special categories (Article 9) or criminal convictions (Article 10)? |
| Consequences | Article 6(4)(d) | What are the possible consequences of the further processing for data subjects? |
| Safeguards | Article 6(4)(e) | Are appropriate safeguards in place, including encryption or pseudonymisation? |
Compatibility is not required when:
- Processing is based on consent (Article 6(1)(a)) — new consent can be obtained for the new purpose
- Processing is based on Union or Member State law that constitutes a necessary and proportionate measure (Article 6(4) final paragraph)
- Processing is for archiving in the public interest, scientific/historical research, or statistics per Article 89(1)
Purpose-Tagged Data Architecture
Design Principles
- Purpose binding at ingestion — Every data record is tagged with its collection purpose at the point of entry into the system
- Purpose-based access control — Access policies are defined per purpose, not per data field alone
- Purpose audit trail — Every access is logged with the purpose under which it was authorized
- Purpose expiry — When a purpose is fulfilled, the associated data enters a retention countdown
Architecture Diagram
┌────────────────────────────────────────────────────────────┐
│ Data Collection Points │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Web Form │ │ API Call │ │ IoT Feed │ │ Partner │ │
│ │ P:ONBRD │ │ P:ANALYT │ │ P:MAINT │ │ P:MARKET │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ └──────────────┼──────────────┼──────────────┘ │
└──────────────────────┼──────────────┼──────────────────────┘
│ │
┌──────────────────────▼──────────────▼──────────────────────┐
│ Purpose Tagging Gateway │
│ ┌─────────────────┐ ┌─────────────────────────────────┐ │
│ │ Purpose Registry │ │ Tag Injection Middleware │ │
│ │ (canonical IDs) │ │ (attaches purpose + timestamp) │ │
│ └─────────────────┘ └─────────────────────────────────┘ │
└──────────────────────────────┬─────────────────────────────┘
│
┌──────────────────────────────▼─────────────────────────────┐
│ Purpose-Partitioned Data Store │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ ONBRD │ │ ANALYT │ │ MAINT │ │ MARKET │ │
│ │ Partition│ │ Partition│ │ Partition│ │ Partition│ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
└──────────────────────────────┬─────────────────────────────┘
│
┌──────────────────────────────▼─────────────────────────────┐
│ Purpose Enforcement Layer │
│ ┌─────────────────┐ ┌────────────────┐ ┌────────────┐ │
│ │ Policy Engine │ │ Compatibility │ │ Audit │ │
│ │ (OPA / Cedar) │ │ Assessor │ │ Logger │ │
│ └─────────────────┘ └────────────────┘ └────────────┘ │
└────────────────────────────────────────────────────────────┘
Purpose Registry Schema
Each processing purpose is registered with canonical metadata:
{
"purpose_id": "PRP-ONBRD-001",
"purpose_name": "Customer Account Onboarding",
"description": "Collection and processing of personal data necessary to create and activate a customer account, verify identity, and establish the contractual relationship.",
"lawful_basis": "Article 6(1)(b)",
"data_categories": ["email", "display_name", "country_code", "identity_verification_result"],
"controller": "Prism Data Systems AG",
"retention_period_days": 2555,
"retention_trigger": "account_closure",
"compatible_purposes": ["PRP-SUPRT-001", "PRP-SECUR-001"],
"incompatible_purposes": ["PRP-MARKET-001", "PRP-RESRCH-001"],
"special_categories": false,
"created_date": "2025-06-01",
"last_reviewed": "2026-01-15",
"review_cadence_days": 180,
"owner": "product_team",
"dpo_approved": true
}
Access Control Per Purpose
Policy-as-Code with Open Policy Agent (OPA)
Purpose-based access control is implemented using OPA policies that evaluate access requests against the purpose registry:
# Prism Data Systems AG — Purpose-Based Access Policy
# Rule: Access is granted only if the requesting service's declared purpose
# matches a purpose tag on the requested data record.
allow {
input.action == "read"
input.resource.purpose_tags[_] == input.requester.declared_purpose
purpose_is_active(input.requester.declared_purpose)
role_authorized_for_purpose(input.requester.role, input.requester.declared_purpose)
}
# Rule: Cross-purpose access requires a completed compatibility assessment
allow {
input.action == "read"
not input.resource.purpose_tags[_] == input.requester.declared_purpose
compatibility_assessment_approved(input.resource.purpose_tags, input.requester.declared_purpose)
purpose_is_active(input.requester.declared_purpose)
}
# Rule: Write operations must include a valid purpose tag
allow {
input.action == "write"
valid_purpose(input.resource.purpose_tag)
input.requester.declared_purpose == input.resource.purpose_tag
}
Role-Purpose Authorization Matrix
| Role | ONBRD | ANALYT | SUPRT | MARKET | SECUR | BILLING |
|---|---|---|---|---|---|---|
| Onboarding Service | Read/Write | — | — | — | — | — |
| Analytics Pipeline | — | Read | — | — | — | — |
| Support Agent | Read | — | Read/Write | — | — | Read |
| Marketing Automation | — | — | — | Read | — | — |
| Security Operations | Read | Read | Read | Read | Read/Write | Read |
| Billing Service | Read | — | — | — | — | Read/Write |
Purpose Compatibility Assessment Workflow
-
Request submission — Service owner submits a request to process data collected under purpose A for new purpose B, documenting the business justification.
-
Factor analysis — The Data Protection Office evaluates the five Article 6(4) factors:
- Link between purposes (scored 1-5: 1 = no link, 5 = closely linked)
- Context and expectations (scored 1-5: 1 = unexpected, 5 = fully expected)
- Nature of data (scored 1-5: 1 = special category, 5 = non-sensitive)
- Consequences (scored 1-5: 1 = severe, 5 = minimal impact)
- Safeguards (scored 1-5: 1 = no safeguards, 5 = comprehensive safeguards)
-
Scoring — Total score calculated (range 5-25):
- 20-25: Compatible. Approve with standard documentation.
- 15-19: Potentially compatible. Approve with additional safeguards (pseudonymization, access restriction).
- 10-14: Likely incompatible. Requires DPO escalation and DPIA consideration.
- 5-9: Incompatible. Denied. New lawful basis or separate consent required.
-
Decision recording — Assessment result recorded in the purpose registry with the assessor identity, date, score breakdown, conditions, and review date.
-
Policy update — If approved, OPA policies are updated to allow cross-purpose access under documented conditions.
Prism Data Systems AG Implementation
Purpose Registry
Prism Data Systems AG maintains 14 registered purposes in their purpose registry:
| Purpose ID | Name | Lawful Basis | Data Categories |
|---|---|---|---|
| PRP-ONBRD-001 | Customer onboarding | Art. 6(1)(b) | email, display_name, country_code |
| PRP-AUTH-001 | Authentication | Art. 6(1)(b) | email, password_hash, mfa_token |
| PRP-BILLING-001 | Billing and invoicing | Art. 6(1)(b) | billing_address, payment_method, vat_id |
| PRP-SUPRT-001 | Customer support | Art. 6(1)(b) | email, display_name, support_ticket_history |
| PRP-ANALYT-001 | Product analytics | Art. 6(1)(f) | pseudonymized_user_id, feature_events, session_duration |
| PRP-MARKET-001 | Direct marketing | Art. 6(1)(a) | email, display_name, marketing_preferences |
| PRP-SECUR-001 | Security monitoring | Art. 6(1)(f) | ip_address, user_agent, login_events |
| PRP-LEGAL-001 | Legal compliance | Art. 6(1)(c) | transaction_records, consent_records |
Enforcement Metrics
Prism Data Systems AG tracks the following purpose limitation enforcement metrics:
| Metric | Target | Measurement |
|---|---|---|
| Purpose tag coverage | 100% of records | Percentage of data records with valid purpose tags |
| Cross-purpose access attempts blocked | > 95% without assessment | Policy engine denial count vs. total cross-purpose requests |
| Compatibility assessments completed | 100% within 5 business days | Time from request to DPO decision |
| Purpose registry review currency | 100% reviewed within cadence | Percentage of purposes reviewed within their review cadence |
Key Regulatory References
- GDPR Article 5(1)(b) — Purpose limitation principle
- GDPR Article 6(4) — Compatibility assessment factors
- GDPR Article 13(1)(c) — Information about purpose at collection
- GDPR Article 14(1)(c) — Information about purpose for indirect collection
- GDPR Article 89(1) — Exception for archiving, research, statistics
- GDPR Recital 50 — Further processing compatibility
- Article 29 Working Party Opinion 03/2013 on purpose limitation (WP203)
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
Conducting LINDDUN Threat Modeling
Overview
LINDDUN is a systematic privacy threat modeling methodology developed by the DistriNet research group at KU Leuven. It provides a structured approach to identifying privacy threats in software systems through Data Flow Diagram (DFD) analysis and threat tree catalogs. LINDDUN stands for seven privacy threat categories:
- Linking — Associating data items with each other or with an individual
- Identifying — Learning the identity of a data subject
- Non-repudiation — Being unable to deny an action or association
- Detecting — Deducing that an individual is involved in a process
- Data Disclosure — Exposing personal data to unauthorized parties
- Unawareness — Data subjects being unaware of data processing
- Non-compliance — Failing to comply with privacy legislation or policies
LINDDUN complements security threat modeling (STRIDE) by focusing specifically on privacy threats. While STRIDE addresses confidentiality, integrity, and availability, LINDDUN addresses unlinkability, anonymity, plausible deniability, undetectability, confidentiality of data content, content awareness, and policy compliance.
LINDDUN Threat Categories Detailed
L — Linking
Definition: An adversary can sufficiently distinguish whether two items of interest (IOI) are related or not within a particular context.
DFD element applicability: Data flows, data stores, processes.
Examples:
- Linking browsing sessions across websites via cookies or fingerprinting
- Linking pseudonymized records across datasets via quasi-identifiers
- Linking transactions to the same individual over time
Threat trees (selected):
- L1: Linking via identifiers shared across contexts
- L2: Linking via quasi-identifier combination (age + postal code + gender)
- L3: Linking via temporal correlation (simultaneous events)
- L4: Linking via behavioral patterns (unique usage signatures)
Mitigations: HIDE (Dissociate, Mix), SEPARATE (Isolate), MINIMIZE (Strip), differential privacy, pseudonymization with context separation.
I — Identifying
Definition: An adversary can sufficiently identify a data subject within a set of data subjects.
DFD element applicability: Data flows, data stores, external entities.
Examples:
- Re-identifying individuals in an anonymized dataset via linkage attacks
- Identifying a user from aggregated location data
- Identifying the author of an anonymous document via stylometry
Threat trees (selected):
- I1: Identification via direct identifiers (name, email, SSN)
- I2: Identification via quasi-identifier combination
- I3: Identification via unique behavioral patterns
- I4: Identification via metadata (IP address, device fingerprint)
Mitigations: HIDE (Encrypt, Obfuscate), MINIMIZE (Strip), ABSTRACT (Group, Summarize), k-anonymity, differential privacy.
N — Non-repudiation
Definition: A data subject is unable to deny having performed an action or being associated with specific data.
DFD element applicability: Data flows, processes, data stores.
Examples:
- Digital signatures on messages prevent denying authorship
- Transaction logs linking purchases to an identity
- Audit trails that irrevocably associate actions with individuals
Threat trees (selected):
- N1: Non-repudiation via digital signatures or cryptographic evidence
- N2: Non-repudiation via witness testimony or log entries
- N3: Non-repudiation via photographic or biometric evidence
Mitigations: HIDE (Mix), group signatures, ring signatures, deniable encryption.
D — Detecting
Definition: An adversary can sufficiently distinguish whether an item of interest (IOI) exists or not.
DFD element applicability: Data flows, processes.
Examples:
- Detecting that an individual visited a specific website via traffic analysis
- Detecting that a patient is in a hospital database (membership inference)
- Detecting the existence of encrypted communication between parties
Threat trees (selected):
- D1: Detection via traffic analysis (communication metadata)
- D2: Detection via timing side channels
- D3: Detection via presence in a dataset (membership inference)
- D4: Detection via access pattern analysis
Mitigations: HIDE (Mix, Obfuscate), cover traffic, onion routing, steganography, ORAM.
D — Data Disclosure
Definition: Personal data is disclosed to or accessed by unauthorized parties.
DFD element applicability: Data flows, data stores, processes.
Examples:
- Unencrypted personal data intercepted in transit
- Database breach exposing customer records
- Employee accessing records without authorization
Threat trees (selected):
- DD1: Disclosure via unencrypted communication channels
- DD2: Disclosure via unauthorized database access
- DD3: Disclosure via side-channel attacks on encrypted data
- DD4: Disclosure via excessive data sharing with third parties
- DD5: Disclosure via improper data deletion (residual data)
Mitigations: HIDE (Encrypt), access control, TLS 1.3, field-level encryption, secure deletion, DLP systems.
U — Unawareness
Definition: Data subjects are unaware of the collection, processing, or sharing of their personal data.
DFD element applicability: External entities (data subjects), processes.
Examples:
- Collecting data without providing a privacy notice
- Processing data for purposes not disclosed to the data subject
- Sharing data with third parties without informing the data subject
Threat trees (selected):
- U1: Unawareness due to missing or inadequate privacy notice
- U2: Unawareness of purpose change (purpose creep without notification)
- U3: Unawareness of third-party sharing
- U4: Unawareness of automated decision-making or profiling
Mitigations: INFORM (Supply, Notify, Explain), layered privacy notices, just-in-time notifications, consent management.
N — Non-compliance
Definition: The system or organization fails to comply with applicable privacy legislation, policies, or standards.
DFD element applicability: All DFD elements.
Examples:
- Processing without a valid lawful basis
- Failing to respond to data subject access requests within 30 days
- Not conducting a required DPIA for high-risk processing
- Retaining data beyond the defined retention period
Threat trees (selected):
- NC1: Non-compliance with lawful basis requirements
- NC2: Non-compliance with data subject rights (Art. 15-22)
- NC3: Non-compliance with cross-border transfer rules (Chapter V)
- NC4: Non-compliance with data breach notification (Art. 33-34)
- NC5: Non-compliance with accountability obligations (Art. 5(2), Art. 24)
Mitigations: ENFORCE (Create, Maintain, Uphold), DEMONSTRATE (Record, Audit, Report), GDPR compliance framework.
LINDDUN Process
Step 1: Define the System Scope
Create a Data Flow Diagram (DFD) of the system showing:
- External entities (data subjects, third parties, regulators)
- Processes (system components that process data)
- Data stores (databases, file systems, logs)
- Data flows (movement of data between elements)
- Trust boundaries (boundaries between different trust domains)
Step 2: Map Threats to DFD Elements
For each DFD element, determine which LINDDUN threat categories apply:
| DFD Element | L | I | N | D | DD | U | NC |
|---|---|---|---|---|---|---|---|
| External entity (data subject) | X | X | |||||
| External entity (third party) | X | X | |||||
| Process | X | X | X | X | X | X | |
| Data store | X | X | X | X | X | ||
| Data flow | X | X | X | X | X | X |
Step 3: Elicit Threats Using Threat Trees
For each applicable threat category on each DFD element, walk through the threat tree catalog to identify specific threats. Document each threat with:
- Threat ID
- Category (LINDDUN letter)
- DFD element
- Description
- Likelihood (1-5)
- Impact (1-5)
- Risk score (likelihood x impact)
Step 4: Prioritize Threats
Rank threats by risk score. Apply risk acceptance thresholds:
- Risk 1-6: Accept with documentation
- Risk 7-12: Mitigate within 6 months
- Risk 13-19: Mitigate within 3 months
- Risk 20-25: Mitigate immediately
Step 5: Select Mitigations
Map each threat to privacy design patterns and specific technical controls. Document the mitigation strategy and responsible team.
Step 6: Validate Mitigations
Verify that selected mitigations adequately address each threat. Update the DFD to reflect implemented controls. Re-assess residual risk.
Key Regulatory References
- GDPR Article 25 — Data protection by design and by default
- GDPR Article 35 — Data protection impact assessment
- GDPR Article 32 — Security of processing
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
- Deng, M., Wuyts, K., Scandariato, R., Preneel, B., & Joosen, W. (2011). "A privacy threat analysis framework: supporting the elicitation and fulfillment of privacy requirements." Requirements Engineering, 16(1), 3-32.
- LINDDUN: linddun.org — Official methodology documentation and threat tree catalogs
Designing Federated Learning Architecture
Overview
Federated Learning (FL) enables multiple parties to collaboratively train a machine learning model without sharing their raw data. Each participant trains a local model on their own data and shares only model updates (gradients or parameters) with a central aggregator. The aggregated model benefits from all participants' data without any single party accessing another's dataset.
FL directly supports GDPR Article 5(1)(c) data minimization (only model updates are shared, not personal data), Article 25(1) data protection by design (privacy is built into the architecture), and can reduce the need for cross-border data transfers under Chapter V (data stays in its jurisdiction of origin).
FL Architectures
Horizontal Federated Learning
Participants share the same feature space but have different data samples. Each participant has a complete record for their subjects but covers different subjects.
┌──────────────────────────────────────────────────────────────┐
│ Central Aggregator │
│ (aggregates model updates only) │
│ │
│ Global Model = Aggregate(Local_1, Local_2, ..., Local_K) │
└─────────────┬──────────────┬──────────────┬──────────────────┘
│ │ │
┌─────────▼────┐ ┌──────▼──────┐ ┌────▼──────────┐
│ Hospital A │ │ Hospital B │ │ Hospital C │
│ 5,000 pts │ │ 8,000 pts │ │ 3,000 pts │
│ Same features│ │ Same features│ │ Same features │
│ Train locally│ │ Train locally│ │ Train locally │
└──────────────┘ └─────────────┘ └───────────────┘
Use case: Multiple hospitals training a diagnostic model; each hospital has complete patient records but different patients.
Vertical Federated Learning
Participants share the same data subjects but have different features. Features are distributed across participants.
┌──────────────────────────────────────────────────────────────┐
│ Coordination Server │
│ (aligns features without exposing raw data) │
└─────────────┬──────────────┬──────────────┬──────────────────┘
│ │ │
┌─────────▼────┐ ┌──────▼──────┐ ┌────▼──────────┐
│ Bank │ │ Insurer │ │ Retailer │
│ Financial │ │ Claims │ │ Purchase │
│ features │ │ features │ │ features │
│ Same users │ │ Same users │ │ Same users │
└──────────────┘ └─────────────┘ └───────────────┘
Use case: Credit scoring where a bank has financial data, an insurer has claims history, and a retailer has purchase behavior — all for overlapping customers.
Aggregation Strategies
FedAvg (Federated Averaging)
The foundational FL algorithm (McMahan et al., 2017). Each participant trains for multiple local epochs, then the aggregator averages the model weights proportional to each participant's dataset size.
For each communication round t:
1. Server sends global model w_t to selected participants
2. Each participant k trains locally for E epochs on their data:
w_k = LocalSGD(w_t, data_k, E epochs, learning_rate η)
3. Server aggregates:
w_{t+1} = Σ (n_k / n) * w_k
where n_k = participant k's sample count, n = total samples
| Parameter | Recommended Range | Impact |
|---|---|---|
| Local epochs (E) | 1-5 | More epochs = less communication, but higher divergence risk |
| Participation rate | 10-100% per round | Higher = better convergence, more communication |
| Learning rate (η) | 0.01-0.1 | Standard SGD tuning applies |
| Communication rounds | 50-500 | Depends on data heterogeneity and model complexity |
FedProx (Federated Proximal)
Extension of FedAvg for heterogeneous settings. Adds a proximal term to the local objective to limit how far local models drift from the global model. Suitable when participants have non-IID (non-independently and identically distributed) data.
Local objective for participant k:
minimize F_k(w) + (μ/2) * ||w - w_t||^2
where μ controls the strength of the proximal regularization
| Scenario | FedAvg Performance | FedProx Performance | Recommendation |
|---|---|---|---|
| IID data across participants | Good | Good | FedAvg (simpler) |
| Mild non-IID (label skew) | Moderate | Good | FedProx (μ=0.01) |
| Severe non-IID (feature shift) | Poor | Moderate | FedProx (μ=0.1) + larger participation |
| Participants with varying compute | Unstable | Stable | FedProx (handles partial work) |
Privacy Guarantees
Threat Model
| Threat | Description | Mitigation |
|---|---|---|
| Gradient inversion | Adversary reconstructs training data from shared gradients | Secure aggregation + differential privacy |
| Membership inference | Adversary determines if a specific record was in training data | Differential privacy |
| Model inversion | Adversary recovers sensitive features from the trained model | Differential privacy + access control on model |
| Free-rider attack | Participant contributes noise instead of genuine updates | Contribution verification (FoolsGold, RFFL) |
| Poisoning attack | Malicious participant corrupts the global model | Byzantine-robust aggregation (Krum, trimmed mean) |
Secure Aggregation
Secure aggregation ensures the server sees only the sum of all participants' updates, not individual updates. Implemented using:
- Pairwise masking (Bonawitz et al., 2017) — Each pair of participants generates a shared random mask. Masks cancel out in the sum, but individual updates remain hidden.
- Threshold secret sharing — Each participant secret-shares their update; the server reconstructs only the aggregate.
- Homomorphic encryption — Participants encrypt updates; server aggregates ciphertexts.
Differential Privacy in FL
Adding calibrated noise to model updates before sharing provides a formal privacy guarantee:
DP-FedAvg:
1. Each participant clips gradients: ||g_k|| ≤ C (sensitivity bounding)
2. Each participant adds Gaussian noise: g_k' = g_k + N(0, σ²·C²·I)
3. Server aggregates noised updates: g_agg = (1/K) Σ g_k'
4. Privacy accounting: track cumulative (ε, δ) via Rényi DP composition
Privacy budget for Prism Data Systems AG FL deployment:
| Parameter | Value | Justification |
|---|---|---|
| Clipping norm (C) | 1.0 | Standard for gradient clipping in DP-SGD |
| Noise multiplier (σ) | 1.1 | Provides (ε=3, δ=1e-5) per training run with 500 rounds |
| Target epsilon (ε) | 3.0 | Moderate privacy for non-special-category data |
| Target delta (δ) | 1e-5 | Standard for n ≈ 316,000 total records |
| Composition method | Rényi DP | Tighter composition bounds than basic composition |
Communication Efficiency
| Technique | Compression Ratio | Accuracy Impact | Description |
|---|---|---|---|
| Gradient quantization | 8-32x | Minimal | Reduce gradient precision from 32-bit to 8-bit or lower |
| Top-k sparsification | 10-100x | Moderate | Transmit only the k largest gradient components |
| Federated dropout | 2-10x | Minimal | Each participant trains a random submodel |
| Gradient compression (SignSGD) | 32x | Moderate | Transmit only the sign of each gradient component |
| Model distillation | Variable | Moderate | Participants share soft predictions instead of gradients |
Implementation Workflow
-
Define the ML task — Specify the model architecture, training objective, and evaluation metrics.
-
Assess data distribution — Evaluate the degree of non-IID across participants. This determines the aggregation strategy (FedAvg for IID, FedProx for non-IID).
-
Select privacy mechanisms — Choose secure aggregation for gradient confidentiality and differential privacy for formal guarantees. Set epsilon budget.
-
Configure communication — Select gradient compression technique based on bandwidth constraints. Set the communication round budget.
-
Implement local training — Deploy the local training pipeline at each participant. Implement gradient clipping and noise injection for DP.
-
Deploy aggregation server — Deploy the central aggregator with secure aggregation protocol. Implement Byzantine-robust aggregation if untrusted participants exist.
-
Monitor and evaluate — Track model convergence, per-participant contribution quality, privacy budget consumption, and communication overhead.
Key Regulatory References
- GDPR Article 5(1)(c) — Data minimization (only model updates shared)
- GDPR Article 25(1) — Data protection by design
- GDPR Article 26 — Joint controller arrangements (FL consortium governance)
- GDPR Chapter V — Cross-border transfers (FL reduces transfer need)
- GDPR Article 35 — DPIA for FL involving special category data
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
Designing Privacy-Preserving Analytics
Overview
Privacy-preserving analytics enables organizations to extract statistical insights from personal data without exposing individual-level information. This directly supports GDPR Article 5(1)(c) (data minimization) and Recital 26 (which exempts truly anonymous data from the regulation). The Article 29 Working Party Opinion 05/2014 on Anonymisation Techniques (WP216) established that effective anonymization must resist singling out, linkability, and inference attacks.
Four primary statistical disclosure control techniques form the foundation of privacy-preserving analytics: differential privacy, k-anonymity, l-diversity, and t-closeness. Each offers different trade-offs between privacy guarantees and data utility.
Statistical Disclosure Control Techniques
Differential Privacy
Differential privacy (Dwork et al., 2006) provides a mathematical guarantee that the output of an analysis is approximately the same whether or not any single individual's data is included. Formally, a randomized mechanism M satisfies (epsilon, delta)-differential privacy if for all datasets D1 and D2 differing in at most one record, and for all sets of outputs S:
P[M(D1) ∈ S] ≤ e^ε × P[M(D2) ∈ S] + δ
Where epsilon (ε) is the privacy loss parameter and delta (δ) bounds the probability of privacy breach.
Privacy Budget Allocation:
| Epsilon Range | Privacy Level | Typical Use Cases |
|---|---|---|
| 0.01 — 0.1 | Very strong | Medical research, genetic data, highly sensitive analytics |
| 0.1 — 1.0 | Strong | General-purpose analytics, demographic analysis |
| 1.0 — 5.0 | Moderate | Aggregate business metrics, trend analysis |
| 5.0 — 10.0 | Weak | Low-sensitivity counts, already-public statistics |
Key Libraries:
| Library | Maintainer | Language | Mechanism Types |
|---|---|---|---|
| Google DP Library | C++/Java/Go | Laplace, Gaussian, partition selection | |
| OpenDP | Harvard IQSS & Microsoft | Rust/Python | Composable framework, Laplace, Gaussian, exponential |
| Apple PPML | Apple | Swift | Local DP, count-mean-sketch, Hadamard response |
| IBM diffprivlib | IBM Research | Python | Scikit-learn compatible, ML with DP |
| PyDP | OpenMined | Python (C++ backend) | Python wrapper around Google DP library |
k-Anonymity
A dataset satisfies k-anonymity (Sweeney, 2002) if every record is indistinguishable from at least k-1 other records with respect to quasi-identifier attributes. Quasi-identifiers are attributes that could be combined with external data to re-identify individuals (e.g., age, postal code, gender).
Implementation approach:
- Identify quasi-identifiers in the dataset
- Apply generalization hierarchies (e.g., age 27 → age range 25-30, postal code 8001 → 80**)
- Apply suppression for records that cannot achieve k-anonymity through generalization alone
- Verify k-value across all equivalence classes
Limitations: k-anonymity does not protect against attribute disclosure when sensitive values within an equivalence class are homogeneous.
l-Diversity
l-Diversity (Machanavajjhala et al., 2007) extends k-anonymity by requiring that each equivalence class contains at least l "well-represented" values of the sensitive attribute. This prevents attribute disclosure attacks.
Variants:
- Distinct l-diversity: Each equivalence class has at least l distinct sensitive values
- Entropy l-diversity: Entropy of sensitive values in each class ≥ log(l)
- Recursive (c,l)-diversity: The most frequent sensitive value appears less than c times the frequency of the least frequent value
t-Closeness
t-Closeness (Li et al., 2007) requires that the distribution of a sensitive attribute in any equivalence class is within distance t of the distribution of the attribute in the entire dataset, measured using Earth Mover's Distance (EMD).
This prevents skewness attacks where an adversary can infer sensitive attributes from the distribution within an equivalence class, even when l-diversity is satisfied.
Privacy Budget Management
Privacy budgets track cumulative privacy loss across multiple queries. Under sequential composition, the total privacy loss is the sum of individual epsilons. Under parallel composition (disjoint subsets), the total privacy loss is the maximum individual epsilon.
Budget Allocation Framework for Prism Data Systems AG:
| Analytics Function | Epsilon Allocation | Refresh Cadence | Justification |
|---|---|---|---|
| Daily active user counts | 0.1 per day | Daily | Low sensitivity, high frequency |
| Revenue by region | 0.5 per quarter | Quarterly | Medium sensitivity, aggregate metric |
| Feature usage patterns | 0.3 per month | Monthly | Used for product development under Art. 6(1)(f) |
| Customer churn analysis | 0.2 per quarter | Quarterly | Involves behavioral profiling |
| A/B test results | 0.1 per experiment | Per experiment | Binary outcome, low disclosure risk |
| Total annual budget | ≤ 8.0 | Sum across all functions with composition |
Budget Exhaustion Protocol:
- When 80% of annual budget is consumed, alert the Data Protection Officer
- When 95% is consumed, require DPO approval for each additional query
- When 100% is consumed, block all further differentially private queries until the next budget period
- Emergency queries require joint approval from DPO and Chief Information Security Officer
Architecture Design
┌──────────────────────────────────────────────────────────┐
│ Analyst Interface │
│ (SQL-like query submission) │
└──────────────────────┬───────────────────────────────────┘
│
┌──────────────────────▼───────────────────────────────────┐
│ Privacy Gateway │
│ ┌─────────────┐ ┌──────────────┐ ┌────────────────┐ │
│ │ Query Parser │ │ Budget Check │ │ Sensitivity │ │
│ │ & Validator │ │ (ε tracker) │ │ Calibration │ │
│ └──────┬──────┘ └──────┬───────┘ └───────┬────────┘ │
│ └────────────────┼──────────────────┘ │
└──────────────────────────┼───────────────────────────────┘
│
┌──────────────────────────▼───────────────────────────────┐
│ Noise Injection Layer │
│ ┌──────────────┐ ┌──────────────┐ ┌───────────────┐ │
│ │ Laplace │ │ Gaussian │ │ Exponential │ │
│ │ Mechanism │ │ Mechanism │ │ Mechanism │ │
│ └──────────────┘ └──────────────┘ └───────────────┘ │
└──────────────────────────┬───────────────────────────────┘
│
┌──────────────────────────▼───────────────────────────────┐
│ Data Processing Layer │
│ ┌─────────────────┐ ┌──────────────────────────────┐ │
│ │ k-Anonymization │ │ Aggregation Engine │ │
│ │ Pre-processing │ │ (min group size: 11) │ │
│ └─────────────────┘ └──────────────────────────────┘ │
└──────────────────────────┬───────────────────────────────┘
│
┌──────────────────────────▼───────────────────────────────┐
│ Encrypted Data Store │
│ (Field-level AES-256-GCM encrypted) │
└──────────────────────────────────────────────────────────┘
Implementation Workflow
-
Classify Analytics Queries — Categorize each analytics use case by sensitivity (direct identifiers accessed, quasi-identifiers combined, sensitive attributes involved) and assign an epsilon budget allocation.
-
Select Mechanism — Choose the appropriate noise mechanism based on query type: Laplace for counting queries, Gaussian for mean/variance queries requiring (ε,δ)-DP, exponential mechanism for selection queries.
-
Calibrate Sensitivity — Determine the global sensitivity of each query function (maximum change in output when one record is added or removed). Use bounded sensitivity where possible by clipping input values.
-
Implement Budget Tracking — Deploy a centralized privacy budget ledger that records every query's epsilon consumption, enforces composition bounds, and blocks queries when the budget is exhausted.
-
Apply Pre-processing — Where differential privacy alone provides insufficient utility, apply k-anonymity as a pre-processing step to reduce the sensitivity of downstream DP queries.
-
Validate Output — Verify that released statistics do not violate minimum group sizes (11 records per Prism Data Systems AG policy) and that confidence intervals are reported alongside noised results.
-
Audit Trail — Log every privacy-preserving query with: timestamp, analyst identity, query hash, epsilon consumed, mechanism used, and cumulative budget remaining.
Key Regulatory References
- GDPR Article 5(1)(c) — Data minimization principle
- GDPR Article 5(1)(e) — Storage limitation
- GDPR Article 25 — Data protection by design and by default
- GDPR Article 89 — Safeguards for processing for scientific/historical research or statistical purposes
- GDPR Recital 26 — Scope of anonymous information
- GDPR Recital 162 — Statistical purposes
- Article 29 Working Party Opinion 05/2014 on Anonymisation Techniques (WP216)
- ENISA Report: Pseudonymisation techniques and best practices (November 2019)
Implementing Data Minimization Architecture
Overview
Data minimization is a core principle of the GDPR under Article 5(1)(c), requiring that personal data be "adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed." Article 25(1) mandates that controllers implement appropriate technical and organisational measures, such as pseudonymisation, designed to implement data-protection principles effectively and to integrate necessary safeguards into the processing.
The European Data Protection Board (EDPB) Guidelines 4/2019 on Article 25 Data Protection by Design and by Default clarify that data minimization applies across four dimensions: the amount of data collected, the extent of processing, the period of storage, and the accessibility of data. ENISA's 2019 report on pseudonymisation techniques provides the technical foundation for implementing these requirements at scale.
Data Minimization Architecture Layers
Layer 1: Collection Minimization
Reduce data at the point of ingestion before it enters backend systems.
Techniques:
| Technique | Description | GDPR Basis | Implementation Complexity |
|---|---|---|---|
| Schema enforcement | Reject fields not explicitly required for the declared purpose | Art. 5(1)(c), Art. 25(1) | Low |
| Client-side filtering | Strip unnecessary fields in the client SDK before transmission | Art. 5(1)(c) | Medium |
| Progressive collection | Request additional fields only when a specific feature is activated | Art. 5(1)(c), Recital 39 | Medium |
| Purpose-gated forms | Display only form fields relevant to the selected service tier | Art. 5(1)(b), Art. 25(2) | Low |
Prism Data Systems AG Implementation:
Prism Data Systems AG deploys an API gateway validation layer that enforces a strict allowlist of fields per endpoint. The customer onboarding endpoint /api/v2/customers accepts only: email, display_name, country_code, and consent_references[]. Fields like date_of_birth, phone_number, and billing_address are collected only when the customer activates the billing module, implementing progressive collection tied to purpose activation.
Layer 2: Processing Minimization
Reduce the identifiability of data during computation.
Pseudonymization (Article 4(5))
Article 4(5) defines pseudonymisation as "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."
ENISA Pseudonymization Techniques (2019 Report):
| Technique | Reversibility | Collision Risk | Suitable For |
|---|---|---|---|
| Counter-based mapping | Reversible with lookup table | None | Customer IDs, transaction references |
| HMAC-SHA256 with secret key | Reversible with key | Negligible (256-bit) | Cross-system linkage where re-identification is needed |
| Format-preserving encryption (FF1/FF3-1) | Reversible with key | None | Structured data (credit card numbers, SSNs) preserving format constraints |
| Tokenization with vault | Reversible with vault access | None | Payment card data (PCI DSS alignment) |
| Keyed hash with salt rotation | Computationally irreversible after rotation | Low | Session-level analytics where longitudinal tracking is unnecessary |
Prism Data Systems AG Implementation: Prism Data Systems AG uses HMAC-SHA256 pseudonymization for all analytics pipelines. Customer identifiers are pseudonymized at the boundary between the transactional database and the analytics data warehouse. The HMAC key is stored in a Hardware Security Module (HSM) managed by the security operations team, physically and logically separated from the analytics infrastructure per ENISA recommended controls.
Anonymization (Recital 26)
Recital 26 states that the principles of data protection should 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." The Article 29 Working Party Opinion 05/2014 on Anonymisation Techniques (WP216) established three risk criteria:
- Singling out — the ability to isolate a record identifying an individual
- Linkability — the ability to link two records relating to the same individual
- Inference — the ability to deduce the value of an attribute from other attributes
Anonymization Techniques:
| Technique | Singling Out | Linkability | Inference | Data Utility |
|---|---|---|---|---|
| k-Anonymity (k=5) | Mitigated | Partially mitigated | Not mitigated | High |
| l-Diversity (l=3) | Mitigated | Mitigated | Partially mitigated | Medium-High |
| t-Closeness (t=0.15) | Mitigated | Mitigated | Mitigated | Medium |
| Differential privacy (epsilon=1.0) | Mitigated | Mitigated | Mitigated | Configurable |
| Data aggregation (min group=11) | Mitigated | Mitigated | Partially mitigated | Low-Medium |
Layer 3: Storage Minimization
Limit how much identifiable data persists at rest.
Field-Level Encryption Architecture:
┌────────────────────────┐
│ Application Layer │
│ (plaintext in memory) │
└──────────┬─────────────┘
│
┌──────────▼─────────────┐
│ Encryption Service │
│ AES-256-GCM per field │
│ Key: KMS / HSM │
└──────────┬─────────────┘
│
┌────────────────────┼────────────────────┐
│ │ │
┌────────▼───────┐ ┌────────▼───────┐ ┌────────▼───────┐
│ email (enc) │ │ name (enc) │ │ country (clr) │
│ DEK-email-v3 │ │ DEK-name-v3 │ │ (not PII) │
└────────────────┘ └────────────────┘ └────────────────┘
Each personally identifiable field is encrypted with a dedicated Data Encryption Key (DEK) wrapped by a Key Encryption Key (KEK) in AWS KMS or Azure Key Vault. This enables selective decryption: analytics queries on country never require decrypting email or name.
Prism Data Systems AG Implementation: Prism Data Systems AG classifies all database columns into four sensitivity tiers:
| Tier | Classification | Encryption | Access Control | Example Fields |
|---|---|---|---|---|
| T1 | Direct identifier | AES-256-GCM, field-level | Named individuals with business justification | email, full_name, national_id |
| T2 | Quasi-identifier | AES-256-GCM, field-level | Role-based, logged | date_of_birth, postal_code, job_title |
| T3 | Sensitive attribute | AES-256-GCM, field-level | Purpose-restricted, dual approval | health_data, financial_score |
| T4 | Non-identifying | Transport encryption (TLS 1.3) | Standard RBAC | country_code, language_preference |
Layer 4: Access Minimization
Restrict who and what systems can access identifiable data.
Data Masking Patterns:
| Pattern | Description | Use Case |
|---|---|---|
| Static masking | Irreversibly replace PII in non-production databases | Development and QA environments |
| Dynamic masking | Apply masking rules at query time based on the requester's role | Customer support dashboards |
| On-the-fly masking | Mask data in transit between microservices | Inter-service API calls where full PII is unnecessary |
| Tokenized views | Database views that return tokens instead of raw values | Reporting layers, third-party integrations |
Prism Data Systems AG Implementation:
Customer support agents at Prism Data Systems AG see dynamically masked data by default: m***l@example.com for email, ***-***-4892 for phone numbers. Only escalation-tier agents can request unmasked access, which requires a ticket reference, is logged in the audit trail, and auto-expires after 30 minutes.
Data Minimization Assessment Matrix
Use this matrix to evaluate each data field against minimization requirements before approving a new processing activity or system design.
| Assessment Criterion | Question | Scoring |
|---|---|---|
| Necessity | Is this field required to fulfill the stated purpose? | 0 = No, 1 = Partially, 2 = Yes |
| Proportionality | Could a less identifying alternative achieve the same result? | 0 = Yes (use alternative), 1 = Partially, 2 = No alternative exists |
| Aggregation potential | Can this field be aggregated or generalized without losing required utility? | 0 = Fully aggregable, 1 = Partially, 2 = Must remain granular |
| Pseudonymization feasibility | Can this field be pseudonymized for this processing purpose? | 0 = Easily pseudonymized, 1 = With effort, 2 = Not feasible |
| Temporal scope | Is this field needed beyond the immediate transaction? | 0 = No (delete after use), 1 = Short retention, 2 = Long retention required |
| Access scope | How many roles need access to the raw value? | 0 = None (mask/encrypt), 1 = Limited roles, 2 = Broad access required |
Scoring interpretation:
- 0-4: Strong candidate for elimination, aggregation, or pseudonymization
- 5-8: Apply masking, field-level encryption, and access controls
- 9-12: Justified retention with full technical safeguards and documentation
Prism Data Systems AG Implementation: Before any new microservice is deployed, the data architecture review board at Prism Data Systems AG requires a completed minimization assessment for every personal data field. Fields scoring below 5 must be eliminated or pseudonymized before the service passes the privacy gate in the CI/CD pipeline.
Key Regulatory References
- GDPR Article 4(5) — Definition of pseudonymisation
- GDPR Article 5(1)(c) — Data minimization principle
- GDPR Article 25(1) — Data protection by design
- GDPR Article 25(2) — Data protection by default
- GDPR Article 32(1)(a) — Pseudonymisation and encryption as security measures
- GDPR Recital 26 — Scope of anonymous information
- GDPR Recital 78 — Appropriate technical and organisational measures
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
- ENISA Report: Pseudonymisation techniques and best practices (November 2019)
- Article 29 Working Party Opinion 05/2014 on Anonymisation Techniques (WP216)
Implementing Data Protection by Default
Overview
GDPR Article 25(2) requires that controllers implement appropriate technical and organisational measures for ensuring that, by default, only personal data which are necessary for each specific purpose of the processing are processed. This obligation applies across four dimensions:
- Amount of personal data collected
- Extent of processing performed
- Period of storage
- Accessibility of personal data
The EDPB Guidelines 4/2019 on Article 25 clarify that "by default" means the strictest privacy settings apply automatically, without requiring any action from the data subject. The data subject should not need to take action to protect their privacy — the system does it for them.
The Four Dimensions of By-Default Protection
Dimension 1: Minimum Data Collection (Amount)
Principle: By default, collect only the data fields strictly necessary for the core service purpose. Additional data collection requires explicit opt-in.
| Default State | Opt-In State | Example |
|---|---|---|
| Core fields only | Extended profile | Account creation collects email + display_name only; phone, address, photo collected only if user activates billing or delivery features |
| No location data | Location-enabled features | Mobile app does not request location permission by default; only requests when user activates location-dependent feature |
| No behavioral tracking | Analytics opt-in | Website loads without analytics cookies; tracking begins only after user grants consent in cookie banner |
| Minimal form fields | Progressive disclosure | Registration form shows 3 required fields; optional fields appear only when user clicks "Add more information" |
Prism Data Systems AG Implementation:
The default customer registration at Prism Data Systems AG collects exactly three fields: email, display_name, and country_code. The country code is needed for regulatory compliance (determining applicable data protection law). All other fields (phone number, billing address, company name, job title) are collected only when the user activates a feature that requires them (billing module, enterprise integration, support escalation).
Dimension 2: Minimum Processing (Extent)
Principle: By default, process data only for the primary purpose. Secondary processing (analytics, profiling, marketing) requires explicit opt-in.
| Processing Activity | Default State | Opt-In Required |
|---|---|---|
| Core service delivery | Active (Art. 6(1)(b)) | No |
| Product analytics | Disabled | Yes — consent toggle in preference center |
| Marketing communications | Disabled | Yes — separate consent at registration |
| Profiling for recommendations | Disabled | Yes — explicit opt-in with explanation |
| Third-party data sharing | Disabled | Yes — named third party + purpose |
| Automated decision-making | Disabled | Yes — Art. 22 explicit consent |
Prism Data Systems AG Implementation: New customer accounts at Prism Data Systems AG have all optional processing activities disabled by default. The account dashboard shows five processing toggles, all set to OFF:
- Product analytics: OFF
- Marketing emails: OFF
- Feature recommendations: OFF
- Usage data for ML training: OFF
- Anonymous benchmarking contributions: OFF
Dimension 3: Minimum Storage (Period)
Principle: By default, apply the shortest defensible retention period. Longer retention requires justification tied to a specific legal obligation.
| Data Category | Default Retention | Extended Retention (Justified) |
|---|---|---|
| Session data | Deleted at session end | Not applicable |
| Server logs | 90 days | 12 months (security investigation) |
| Customer account data | Account duration + 30 days | + statutory period (tax: 10 years for billing) |
| Support tickets | 1 year after closure | 3 years (limitation period) |
| Marketing consent records | Consent duration + 3 years | + 5 years (accountability, Art. 7(1)) |
| Analytics data | 13 months | Not applicable (maximum) |
Prism Data Systems AG Implementation: The retention policy engine at Prism Data Systems AG applies the shortest retention period by default for each data category. Extended retention is applied only for data categories with documented statutory requirements (tax records: 10 years per Swiss CO Art. 958f). All other data follows the purpose-based minimum retention.
Dimension 4: Minimum Accessibility
Principle: By default, restrict access to personal data to the minimum set of roles necessary. Broader access requires explicit authorization.
| Role | Default Access | Elevated Access (Requires Justification) |
|---|---|---|
| Customer support (Tier 1) | Masked PII only (m***@domain.ch) | Unmasked PII (30-min timeout, ticket reference required) |
| Analytics team | Pseudonymized data only | No access to direct identifiers |
| Engineering team | No production PII access | Break-glass access (dual approval, logged, time-limited) |
| Marketing team | Aggregate statistics only | No individual-level data access |
| DPO | Full access for compliance investigations | Standing access with audit trail |
| External auditor | Anonymized samples | Purpose-limited access with NDA |
Prism Data Systems AG Implementation: All new employee accounts at Prism Data Systems AG are provisioned with zero personal data access by default. Access is granted per-purpose through an access request workflow requiring manager approval and DPO notification. Access grants expire after 90 days and must be renewed. The Tier 1 support dashboard displays dynamically masked data by default.
Technical Implementation Checklist
User Interface Defaults
- All consent toggles default to OFF (unticked)
- Privacy settings page accessible from main navigation
- No pre-selected optional data collection
- No dark patterns (no misleading button labels, no visual emphasis on "Accept All")
- Cookie banner defaults to "Reject All" with equal visual prominence to "Accept All"
- Registration form shows only mandatory fields by default
- Notification preferences default to minimum (transactional only)
- Profile visibility defaults to private (not public)
- Data sharing toggles default to OFF for each third party
- Search engine indexing of profiles defaults to OFF
Backend Defaults
- API endpoints reject non-allowlisted fields (return 400 for unexpected fields)
- New database tables include TTL metadata columns by default
- Default retention period applied automatically unless longer period is explicitly configured
- Pseudonymization applied at analytics ingestion boundary by default
- Field-level encryption enabled for all PII columns by default
- Dynamic masking rules applied to all support and operations dashboards
- Audit logging enabled for all personal data access by default
- Minimum group size (11) enforced on all analytics outputs
- Cross-purpose data access blocked by default (requires compatibility assessment)
- New service deployments fail privacy gate if data minimization assessment is not completed
Access Control Defaults
- New employee accounts provisioned with zero personal data access
- Access grants expire after 90 days by default
- Elevated access requires documented justification and approval
- Break-glass access logged and reviewed within 24 hours
- Third-party access disabled by default until DPA is executed
- API keys scoped to minimum required data categories
Communication Defaults
- Marketing communications disabled for all new accounts
- Transactional emails only (account confirmations, security alerts)
- Push notifications limited to security alerts by default
- Newsletter subscription requires separate opt-in
- Partner communications disabled by default
By-Default Verification Process
-
New account audit — Create a test account through the standard registration flow. Verify that no optional processing is active, no marketing communications are scheduled, and the minimum data fields are collected.
-
Dashboard inspection — Log in as a Tier 1 support agent. Verify that all PII is masked by default and that unmasking requires explicit action with justification.
-
API testing — Send API requests with additional fields beyond the allowlist. Verify that extra fields are rejected (HTTP 400).
-
Retention verification — Check that newly created records have TTL metadata set according to the minimum retention period.
-
Analytics output — Run a standard analytics query. Verify that results are pseudonymized, minimum group size is enforced, and differential privacy noise is applied.
Key Regulatory References
- GDPR Article 25(2) — Data protection by default
- GDPR Article 25(1) — Data protection by design
- GDPR Recital 78 — Technical measures for data protection by default
- EDPB Guidelines 4/2019 on Article 25, Section 3: Data Protection by Default
- CJEU C-673/17 (Planet49) — Pre-ticked boxes do not constitute consent
- EDPB Guidelines 05/2020 on consent — Consent must be freely given, not default
Implementing Homomorphic Encryption
Overview
Homomorphic encryption (HE) enables computation on encrypted data without decryption. The data owner encrypts their data, sends the ciphertext to a computing party (e.g., a cloud processor under Article 28), and the computing party performs operations on the ciphertext. The encrypted results are returned to the data owner, who decrypts them to obtain the plaintext result. At no point does the computing party see unencrypted data.
This directly supports GDPR Article 32(1)(a), which identifies encryption as an appropriate technical measure, and Article 25(1), which requires data protection by design. HE enables a controller to outsource computation to a processor while maintaining cryptographic confidentiality of personal data.
HE Scheme Selection
Scheme Comparison
| Scheme | Arithmetic Type | Exact/Approximate | Operations | Best For | Libraries |
|---|---|---|---|---|---|
| BFV | Integer (modular) | Exact | Add, Multiply (limited depth) | Counting, matching, exact integer computation | Microsoft SEAL, OpenFHE |
| BGV | Integer (modular) | Exact | Add, Multiply (with modulus switching) | Deep circuits on integers, batched operations | HELib, OpenFHE |
| CKKS | Real/Complex | Approximate | Add, Multiply, Rotation | ML inference, statistical analysis, floating-point computation | Microsoft SEAL, OpenFHE, HEAAN |
| TFHE | Boolean/Integer | Exact | Arbitrary (gate-by-gate) | Programmable bootstrapping, arbitrary functions | TFHE-rs, Concrete (Zama) |
Decision Criteria
| Criterion | BFV/BGV | CKKS | TFHE |
|---|---|---|---|
| Need exact results? | Yes | No (approximate) | Yes |
| Data type | Integers | Real numbers | Boolean/Small integers |
| Multiplicative depth | Limited (plan ahead) | Limited (plan ahead) | Unlimited (with bootstrapping) |
| SIMD batching | Yes (thousands of slots) | Yes (thousands of slots) | Limited |
| Performance | Fast for shallow circuits | Fast for ML workloads | Slow per operation, fast bootstrapping |
| Typical latency | Milliseconds (shallow) to seconds | Milliseconds to seconds | Seconds to minutes |
Parameter Selection Guide
For BFV/CKKS with Microsoft SEAL, the key parameters are:
| Parameter | Description | Security Impact | Performance Impact |
|---|---|---|---|
| poly_modulus_degree (N) | Degree of the polynomial ring | Higher N = more security | Higher N = slower operations, more memory |
| coeff_modulus | Product of primes defining the coefficient modulus | Determines noise budget and multiplicative depth | More primes = deeper circuits but larger ciphertexts |
| plain_modulus (BFV only) | Modulus for plaintext encoding | Must accommodate plaintext values | Larger = less noise budget available |
| scale (CKKS only) | Encoding scale factor | Determines precision of approximate results | Higher scale = more precision but faster noise growth |
Recommended parameter sets (128-bit security):
| Use Case | N | coeff_modulus bits | Mult Depth | Memory/Ciphertext |
|---|---|---|---|---|
| Simple addition/comparison | 4096 | [40, 40] | 1 | ~64 KB |
| Moderate computation (sum + multiply) | 8192 | [60, 40, 40, 60] | 2-3 | ~256 KB |
| ML inference (logistic regression) | 16384 | [60, 40, 40, 40, 40, 60] | 4-5 | ~1 MB |
| Deep neural network inference | 32768 | [60, 40×8, 60] | 8+ | ~4 MB |
Architecture for GDPR-Compliant HE Processing
┌────────────────────────────────────────────────────────────┐
│ Data Controller (Prism Data Systems AG) │
│ ┌──────────────────┐ ┌──────────────────────────────┐ │
│ │ Key Generator │ │ Encrypt personal data │ │
│ │ (sk stays here) │ │ with public key │ │
│ └────────┬─────────┘ └──────────┬───────────────────┘ │
│ │ pk, rlk, gk │ ciphertext │
└───────────┼───────────────────────┼────────────────────────┘
│ │
▼ ▼
┌────────────────────────────────────────────────────────────┐
│ Data Processor (Cloud Service) │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ Compute on ciphertext using pk, rlk, gk │ │
│ │ (NEVER has access to sk — cannot decrypt) │ │
│ │ │ │
│ │ Operations: HE.Add, HE.Multiply, HE.Rotate │ │
│ └──────────────────────────┬───────────────────────────┘ │
│ │ encrypted result │
└──────────────────────────────┼─────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────┐
│ Data Controller (Prism Data Systems AG) │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ Decrypt result using sk │ │
│ │ Plaintext result = f(data) computed without exposure │ │
│ └──────────────────────────────────────────────────────┘ │
└────────────────────────────────────────────────────────────┘
Key:
sk = Secret key (never leaves controller)
pk = Public key (for encryption and computation)
rlk = Relinearization keys (for ciphertext multiplication)
gk = Galois keys (for ciphertext rotation/SIMD operations)
Performance Considerations
Benchmark Reference (Microsoft SEAL, BFV, N=8192, 128-bit security)
| Operation | Latency | Throughput (ops/sec) |
|---|---|---|
| Key generation | 50 ms | 20 |
| Encryption | 2 ms | 500 |
| Decryption | 1 ms | 1,000 |
| Ciphertext addition | 0.1 ms | 10,000 |
| Ciphertext multiplication | 5 ms | 200 |
| Relinearization | 3 ms | 333 |
| SIMD rotation | 4 ms | 250 |
Optimization Strategies
| Strategy | Benefit | Trade-off |
|---|---|---|
| SIMD batching | Process thousands of values in parallel | Requires data to be structured as vectors |
| Circuit depth minimization | Fewer multiplications = less noise, smaller parameters | May require algorithmic redesign |
| Lazy relinearization | Skip intermediate relinearizations | Larger intermediate ciphertexts |
| Level-aware computation | Use modulus switching to control noise | Requires careful depth planning |
| Hybrid HE + plaintext | Perform non-sensitive operations in plaintext | Reduces privacy guarantee scope |
Implementation Workflow
-
Identify computation — Define the exact mathematical function to compute on encrypted data. Express it as additions and multiplications (and rotations for SIMD).
-
Select scheme — Choose BFV for exact integer results, CKKS for approximate real-number results, TFHE for arbitrary boolean/integer functions.
-
Estimate circuit depth — Count the maximum number of sequential multiplications in the computation graph. This determines parameter requirements.
-
Select parameters — Choose poly_modulus_degree and coeff_modulus to support the required depth at 128-bit security.
-
Implement and test — Implement the computation using an HE library. Test against plaintext reference implementation to verify correctness.
-
Benchmark — Measure encryption, computation, and decryption latency. Verify memory requirements are within infrastructure limits.
-
Deploy — Establish key management (secret key in HSM at controller, public keys distributed to processor), monitoring, and audit logging.
Key Regulatory References
- GDPR Article 25(1) — Data protection by design (encryption as a design measure)
- GDPR Article 32(1)(a) — Encryption as a security measure
- GDPR Article 28 — Processor obligations (HE enables processing without plaintext exposure)
- GDPR Recital 83 — Encryption to maintain security
- ENISA Report: Data Protection Engineering (2022)
- HomomorphicEncryption.org Standard: Homomorphic Encryption Standardization (2018)
Implementing Secure Multi-Party Computation
Overview
Secure Multi-Party Computation (SMPC) enables multiple parties to jointly compute a function over their combined inputs while keeping each party's individual input private. No party learns anything beyond the output of the computation and what can be inferred from their own input and the output.
SMPC supports GDPR Article 5(1)(c) data minimization by eliminating the need to centralize data, Article 25(1) data protection by design by building privacy into the computation architecture, and Article 26 joint controller arrangements by enabling collaborative analytics without data sharing.
Core SMPC Techniques
Shamir Secret Sharing
Shamir's Secret Sharing (1979) splits a secret value into n shares such that any t shares can reconstruct the secret (threshold t-out-of-n), but fewer than t shares reveal no information.
Properties:
- Information-theoretic security (unbreakable regardless of computational power)
- Supports addition of shared values without communication
- Multiplication requires an interactive protocol (Beaver triples or resharing)
Application: Splitting personal data across multiple servers such that no single server (or coalition below threshold) can reconstruct the data.
Garbled Circuits
Yao's Garbled Circuits (1986) enable two-party computation. One party (the garbler) encrypts a boolean circuit; the other party (the evaluator) evaluates the encrypted circuit without learning intermediate values.
Properties:
- Constant-round protocol (efficient for high-latency networks)
- Communication cost proportional to circuit size
- Best for two-party computation with complex boolean functions
Oblivious Transfer
Oblivious Transfer (OT) is a protocol where a sender has multiple messages and a receiver selects one message to receive, without the sender learning which message was selected and without the receiver learning the other messages.
Role in SMPC: OT is the foundational building block for garbled circuit evaluation and is used for input wire labels in Yao's protocol.
SMPC Frameworks
| Framework | Protocol | Language | Security Model | Best For |
|---|---|---|---|---|
| MP-SPDZ | Multiple (SPDZ, MASCOT, semi-honest, malicious) | Python-like DSL | Semi-honest and malicious | Research and prototyping with multiple security models |
| CrypTen | Secret sharing (2-party and 3-party) | Python (PyTorch) | Semi-honest | ML inference and training on shared data |
| MOTION | GMW, BMR, arithmetic/boolean sharing | C++ | Semi-honest | High-performance 2+ party computation |
| ABY/ABY3 | Arithmetic, Boolean, Yao sharing | C++ | Semi-honest (2-party ABY, 3-party ABY3) | Mixed-protocol computation (switching between share types) |
| Sharemind | Additive secret sharing (3-party) | SecreC (DSL) | Semi-honest (honest majority) | Enterprise deployment with managed infrastructure |
MP-SPDZ Protocol Selection
| Protocol | Security Model | Parties | Preprocessing | Online Performance |
|---|---|---|---|---|
| SPDZ/MASCOT | Malicious (dishonest majority) | 2+ | Heavy (OT-based) | Fast online phase |
| Semi2k | Semi-honest (dishonest majority) | 2+ | Moderate | Fast |
| Shamir | Semi-honest (honest majority) | 3+ | Light | Very fast |
| Rep3 | Semi-honest (honest majority) | 3 | None | Very fast |
| MASCOT | Malicious (dishonest majority) | 2+ | OT-based | Moderate |
| Yao's GC | Semi-honest | 2 | Garbling | Fast evaluation |
Architecture for GDPR-Compliant SMPC
┌─────────────────────────────────────────────────────────────┐
│ SMPC Computation Coordinator │
│ (orchestrates protocol execution) │
└──────┬─────────────────┬─────────────────┬──────────────────┘
│ │ │
┌──────▼──────┐ ┌───────▼──────┐ ┌──────▼──────┐
│ Party A │ │ Party B │ │ Party C │
│ (Bank) │ │ (Insurer) │ │ (Retailer) │
│ │ │ │ │ │
│ Input: x_A │ │ Input: x_B │ │ Input: x_C │
│ Share: [x_A]│ │ Share: [x_B] │ │ Share: [x_C]│
│ │ │ │ │ │
│ Compute on │ │ Compute on │ │ Compute on │
│ local share │ │ local share │ │ local share │
└──────┬──────┘ └───────┬──────┘ └──────┬──────┘
│ │ │
└────────────────►│◄────────────────┘
│
┌──────▼──────┐
│ Output: │
│ f(x_A,x_B, │
│ x_C) │
│ (revealed │
│ to all) │
└─────────────┘
Each party learns ONLY the final output f(x_A, x_B, x_C),
NOT the individual inputs of other parties.
Practical Deployment Patterns
Pattern 1: Private Set Intersection (PSI)
Two parties determine which records they have in common without revealing records unique to either party.
Use case: A bank and an insurer identify shared customers for a joint risk assessment without revealing their full customer lists.
| Step | Action | Privacy Guarantee |
|---|---|---|
| 1 | Each party hashes their customer identifiers with a shared key | Individual identifiers not revealed |
| 2 | Parties engage in PSI protocol (DH-based or OT-based) | Only intersection elements revealed |
| 3 | For matched records, proceed with joint computation on shared data | Non-matched records remain private |
Pattern 2: Private Aggregation
Multiple parties compute aggregate statistics (sum, mean, count) over their combined data without revealing individual contributions.
Use case: Three regional offices of Prism Data Systems AG compute total headcount and average salary without revealing per-office figures.
Pattern 3: Private ML Training
Multiple parties jointly train a machine learning model using SMPC to protect training data during the process.
Use case: Using CrypTen, two hospitals jointly train a logistic regression model on encrypted patient features.
Performance Considerations
| Factor | Impact | Mitigation |
|---|---|---|
| Network latency | SMPC is communication-intensive; high latency increases total time | Co-locate computation nodes; use constant-round protocols (Yao) for WAN |
| Data size | Communication scales with input size | Pre-aggregate locally where possible; use sketching |
| Circuit complexity | Deeper circuits require more communication rounds (for GMW) | Optimize circuit depth; use Yao for constant rounds |
| Number of parties | More parties = more communication pairs | Use 3-party protocols with honest majority for efficiency |
| Security model | Malicious security is 10-100x slower than semi-honest | Use semi-honest for trusted consortium members; reserve malicious for adversarial settings |
Implementation Workflow
- Define the computation — Express the joint function as arithmetic or boolean operations.
- Select security model — Semi-honest if all parties are trusted (e.g., consortium members); malicious if adversarial behavior is possible.
- Select framework — Choose MP-SPDZ for flexibility, CrypTen for ML workloads, Sharemind for managed enterprise.
- Implement protocol — Write the computation in the framework's DSL, test on synthetic data.
- Benchmark — Measure end-to-end latency, communication volume, and memory usage.
- Deploy — Establish secure communication channels (mTLS), key distribution, and monitoring.
- Govern — Execute joint controller agreement (Art. 26), update Article 30 records, document in DPIA.
Key Regulatory References
- GDPR Article 5(1)(c) — Data minimization (no data centralization needed)
- GDPR Article 25(1) — Data protection by design
- GDPR Article 26 — Joint controller arrangements for SMPC consortia
- GDPR Article 28 — Processor obligations (SMPC coordinator role)
- GDPR Article 32(1) — Security measures
- GDPR Recital 78 — Technical measures for data protection
- ENISA Report: Data Protection Engineering (2022)
Preparing ISO 31700 Certification
Overview
ISO 31700:2023 "Consumer protection — Privacy by design for consumer goods and services" is the first international standard specifically dedicated to privacy by design. Published in February 2023, it provides 30 requirements organized across three lifecycle phases: design, production/deployment, and end-of-life/disposal.
ISO 31700 operationalizes the privacy by design concept originally proposed by Ann Cavoukian and codified in GDPR Article 25(1). While Article 25 establishes the legal obligation, ISO 31700 provides a certifiable framework for demonstrating compliance through structured requirements across the entire product lifecycle.
ISO 31700 Structure
Part 1: High-Level Requirements (ISO 31700-1:2023)
30 requirements organized into three lifecycle phases.
Part 2: Use Cases (ISO 31700-2:2023)
Practical use cases illustrating application of Part 1 requirements across industries including e-commerce, IoT devices, mobile applications, and cloud services.
The 30 Requirements
Phase 1: Design (Requirements 1-16)
| Req # | Title | Summary | GDPR Mapping |
|---|---|---|---|
| 1 | Designing privacy controls | Consumer goods shall include privacy controls that enable consumers to manage collection, use, and disclosure of their PII | Art. 25(1), Art. 25(2) |
| 2 | Privacy notice and consent | Privacy information shall be communicated to consumers in a clear and understandable manner | Art. 12-14 |
| 3 | Consumer PII collection | PII collection shall be limited to what is necessary and proportionate | Art. 5(1)(c) |
| 4 | Creating and managing consumer PII | Processes shall be established for the creation, maintenance, and deletion of PII | Art. 5(1)(d), Art. 5(1)(e) |
| 5 | Consumer PII use limitation | PII shall be processed only for specified, explicit, and legitimate purposes | Art. 5(1)(b) |
| 6 | Consumer PII disclosure to third parties | PII disclosure to third parties shall be limited and controlled | Art. 28, Art. 44-49 |
| 7 | Consumer PII quality | Measures shall ensure accuracy and currency of PII | Art. 5(1)(d) |
| 8 | Consumer PII security | Appropriate security measures shall protect PII | Art. 32 |
| 9 | Consumer PII retention | PII shall not be retained longer than necessary | Art. 5(1)(e) |
| 10 | Organizational commitment | Management shall commit resources and oversight for privacy | Art. 24 |
| 11 | Privacy risk assessment | Privacy risks shall be identified, assessed, and treated | Art. 35 |
| 12 | Designing PII processing rights | Consumer rights to access, correct, delete, and port their PII shall be built into the design | Art. 15-20 |
| 13 | Designing ergonomic privacy controls | Privacy controls shall be usable and accessible to consumers | Art. 25(2) |
| 14 | Designing privacy for vulnerable consumers | Design shall consider the needs of children, elderly, and other vulnerable populations | Art. 8, Recital 38 |
| 15 | Supply chain management | Privacy requirements shall be imposed on suppliers and sub-processors | Art. 28 |
| 16 | Designing for disposal | Design shall include mechanisms for PII deletion at end of product life | Art. 17 |
Phase 2: Production and Deployment (Requirements 17-24)
| Req # | Title | Summary | GDPR Mapping |
|---|---|---|---|
| 17 | Privacy awareness and training | Personnel shall receive privacy training appropriate to their role | Art. 39(1)(b) |
| 18 | Privacy operations | Operational procedures shall implement designed privacy controls | Art. 24, Art. 25(1) |
| 19 | Privacy breach management | Incident response procedures for privacy breaches shall be established | Art. 33, Art. 34 |
| 20 | Communication with consumers | Channels for consumer privacy inquiries and complaints shall be maintained | Art. 12(2) |
| 21 | Third-party management | Third parties processing PII shall be managed and monitored | Art. 28 |
| 22 | Consumer PII use | PII processing in production shall align with designed purposes and controls | Art. 5(1)(b) |
| 23 | Managing privacy inquiries and complaints | Consumer inquiries and complaints shall be handled within specified timeframes | Art. 12(3) |
| 24 | Privacy performance monitoring | Privacy controls shall be monitored for effectiveness | Art. 24(1) |
Phase 3: End-of-Life and Disposal (Requirements 25-30)
| Req # | Title | Summary | GDPR Mapping |
|---|---|---|---|
| 25 | End-of-life PII management | PII shall be managed when a product reaches end of life | Art. 17 |
| 26 | Consumer notification of disposal | Consumers shall be informed about PII handling during disposal | Art. 13 |
| 27 | PII disposal | PII shall be securely deleted or anonymized at end of life | Art. 5(1)(e), Art. 17 |
| 28 | Data portability at end of life | Consumers shall be able to retrieve their data before disposal | Art. 20 |
| 29 | Supply chain disposal | Suppliers and sub-processors shall delete PII upon product disposal | Art. 28(3)(g) |
| 30 | Disposal documentation | Disposal activities shall be documented and verifiable | Art. 5(2) |
Gap Assessment Methodology
Assessment Scoring
For each of the 30 requirements, assess the current state:
| Score | Level | Description |
|---|---|---|
| 0 | Non-existent | No awareness of the requirement |
| 1 | Initial | Awareness exists but no formal implementation |
| 2 | Developing | Partial implementation, ad hoc processes |
| 3 | Defined | Formal processes documented and communicated |
| 4 | Managed | Processes monitored, measured, and consistently applied |
| 5 | Optimizing | Continuous improvement based on metrics and feedback |
Certification readiness: All requirements must be at Level 3 or above. Level 4 or above is recommended for high-risk processing.
Gap Assessment Process
- Document current state — For each requirement, document what controls, processes, and evidence currently exist.
- Score maturity — Apply the 0-5 scoring scale based on documented evidence.
- Identify gaps — Requirements scoring below 3 are certification gaps.
- Prioritize remediation — Rank gaps by: risk impact (high-risk processing first), dependency (foundational requirements first), and effort.
- Create remediation plan — Define specific actions, responsible parties, and target dates for each gap.
- Execute and verify — Implement remediations and verify through evidence collection.
- Pre-certification audit — Conduct internal audit using ISO 31700 checklist before engaging certification body.
Prism Data Systems AG Gap Assessment Summary
| Phase | Requirements | Avg Score | Gaps (< 3) | Critical Gaps (< 2) |
|---|---|---|---|---|
| Design (1-16) | 16 | 3.8 | 2 | 0 |
| Production (17-24) | 8 | 3.5 | 1 | 0 |
| Disposal (25-30) | 6 | 2.7 | 3 | 1 |
| Total | 30 | 3.4 | 6 | 1 |
Critical gap: Requirement 27 (PII disposal) scored 1 — while Prism Data Systems AG has automated deletion for database records, physical media disposal procedures for decommissioned hardware lack documented verification.
Remediation Roadmap
| Priority | Req # | Gap Description | Remediation Action | Owner | Target |
|---|---|---|---|---|---|
| 1 | 27 | Physical media disposal lacks verification | Implement NIST SP 800-88 media sanitization with certificate of destruction | IT Operations | 2026-05-01 |
| 2 | 25 | End-of-life PII management incomplete | Define end-of-life procedures for each product line; assign lifecycle owners | Product Management | 2026-05-15 |
| 3 | 29 | Supplier disposal obligations not enforced | Add disposal audit clause to supplier contracts; annual verification | Procurement | 2026-06-01 |
| 4 | 14 | Vulnerable consumer design considerations informal | Document accessibility and child safety review as design stage gate | UX Engineering | 2026-06-15 |
| 5 | 16 | Disposal design mechanisms incomplete | Add data wipe functionality to all IoT device firmware | Firmware Engineering | 2026-07-01 |
| 6 | 24 | Privacy monitoring metrics not formalized | Define KPIs for each privacy control; implement dashboard | Privacy Engineering | 2026-06-01 |
Implementation Workflow
-
Executive sponsorship — Secure management commitment (Requirement 10) with budget allocation and executive sponsor assignment.
-
Gap assessment — Conduct full assessment of all 30 requirements using the scoring methodology above.
-
Remediation planning — Create prioritized remediation roadmap addressing all gaps below Level 3.
-
Implementation — Execute remediations in priority order, documenting evidence of implementation for each requirement.
-
Internal audit — Conduct pre-certification internal audit using the ISO 31700 checklist to verify all requirements are at Level 3+.
-
Certification body engagement — Select an accredited certification body and schedule the Stage 1 (documentation review) and Stage 2 (implementation audit) assessments.
-
Continuous improvement — After certification, maintain a continuous improvement cycle with annual surveillance audits and periodic internal reviews.
Key Regulatory References
- ISO 31700-1:2023 — Consumer protection — Privacy by design for consumer goods and services — Part 1: High-level requirements
- ISO 31700-2:2023 — Part 2: Use cases
- GDPR Article 25 — Data protection by design and by default
- GDPR Article 42 — Certification
- GDPR Article 43 — Certification bodies
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
- NIST SP 800-88 Rev. 1 — Guidelines for Media Sanitization
Assessing Pseudonymization and Re-Identification Risk
Overview
Pseudonymization (GDPR Article 4(5)) means processing personal data in such a manner that the 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 non-attribution. Unlike anonymization (Recital 26), pseudonymized data remains personal data — but it significantly reduces risk and is explicitly recognized as an appropriate safeguard under Articles 25, 32, and 89.
The ENISA report "Pseudonymisation Techniques and Best Practices" (November 2019) provides a comprehensive taxonomy of pseudonymization techniques and their properties. This skill implements a structured assessment framework for selecting appropriate techniques and quantifying residual re-identification risk.
Pseudonymization Techniques Taxonomy
Technique 1: Counter-Based Pseudonyms
| Property | Value |
|---|---|
| Mechanism | Sequential or random counter assigned to each identifier |
| Reversibility | Reversible via lookup table |
| Collision resistance | Guaranteed (bijective mapping) |
| Performance | O(1) lookup |
| Storage overhead | Mapping table size proportional to identifier count |
| Key management | Protect the mapping table as additional information |
Prism Data Systems AG Implementation: Customer IDs are replaced with UUIDv4 pseudonyms at the analytics ingestion boundary. The mapping table is stored in a separate HSM-backed database accessible only to the DPO and the identity resolution service (which requires dual-approval access).
Technique 2: Cryptographic Hashing (HMAC)
| Property | Value |
|---|---|
| Mechanism | HMAC-SHA256 with secret key applied to identifier |
| Reversibility | Not directly reversible; vulnerable to brute-force on low-entropy inputs |
| Collision resistance | Negligible collision probability (256-bit output) |
| Performance | O(1) computation |
| Storage overhead | Fixed 32-byte output per identifier |
| Key management | HMAC key must be protected as additional information |
Risk: Plain SHA-256 hashing (without key) is vulnerable to rainbow table attacks on low-entropy identifiers like email addresses. HMAC with a secret key mitigates this but the key becomes the single point of re-identification.
Prism Data Systems AG Implementation: HMAC-SHA256 with a 256-bit key rotated quarterly is used for pseudonymizing user identifiers in the analytics pipeline. The HMAC key is stored in AWS KMS with access restricted to the pseudonymization service account. Key rotation triggers re-pseudonymization of the active analytics dataset.
Technique 3: Encryption-Based Pseudonymization
| Property | Value |
|---|---|
| Mechanism | AES-256-GCM or format-preserving encryption (FF1/FF3-1) |
| Reversibility | Fully reversible with decryption key |
| Collision resistance | Guaranteed (bijective for same key) |
| Performance | O(1) computation; FPE slower than standard AES |
| Storage overhead | Ciphertext size ≈ plaintext + 16-byte tag (GCM) |
| Key management | Encryption key is the additional information per Art. 4(5) |
Format-Preserving Encryption preserves the format of the original identifier (e.g., encrypting a 10-digit phone number produces another 10-digit number), enabling pseudonymization without schema changes.
Prism Data Systems AG Implementation: FPE (FF3-1 mode) is used for pseudonymizing phone numbers and IBAN fields in the support ticket system. This allows Tier 1 support agents to see structurally valid but pseudonymized identifiers, reducing accidental exposure while maintaining workflow compatibility.
Technique 4: Tokenization
| Property | Value |
|---|---|
| Mechanism | Random token generated and stored in a secure vault; no mathematical relationship to original |
| Reversibility | Reversible only via vault lookup |
| Collision resistance | Guaranteed by vault uniqueness constraint |
| Performance | O(1) vault lookup; network latency for centralized vault |
| Storage overhead | Vault stores original-to-token mapping |
| Key management | Vault access controls replace key management |
Prism Data Systems AG Implementation: Payment card data is tokenized using a PCI DSS-compliant token vault. The vault is operated by the payment processor and is logically separated from Prism's infrastructure. Detokenization requires a PCI-scoped service account with transaction-level authorization.
Technique 5: Synthetic Data Replacement
| Property | Value |
|---|---|
| Mechanism | Replace real identifiers with structurally similar but fictional values |
| Reversibility | Not reversible (one-way replacement) |
| Collision resistance | Depends on generation method; risk of accidental collision with real data |
| Performance | O(1) generation |
| Storage overhead | No mapping table needed (irreversible) |
| Key management | Not applicable |
Use case: Test environments where referential integrity must be maintained but no re-identification path is acceptable.
Technique Comparison Matrix
| Criterion | Counter | HMAC | Encryption | Tokenization | Synthetic |
|---|---|---|---|---|---|
| Reversibility | Lookup table | Brute-force only | Decryption key | Vault lookup | Irreversible |
| Deterministic | Configurable | Yes (same key) | Yes (same key+IV) | Configurable | No |
| Format preserving | No | No | FPE only | Configurable | Yes |
| Linkability (same dataset) | Yes | Yes | Yes | Configurable | No |
| Linkability (cross-dataset) | No | Same key: Yes | Same key: Yes | No | No |
| Brute-force resistance | N/A | Key-dependent | Key-dependent | N/A | N/A |
| Suited for analytics | Yes | Yes | Limited | Yes | Limited |
| Suited for testing | No | No | No | Yes | Yes |
| ENISA recommendation level | Basic | Intermediate | Advanced | Advanced | Specialized |
Re-Identification Risk Assessment Framework
The Motivated Intruder Test
The UK Information Commissioner's Office (ICO) Anonymisation Code of Practice defines the motivated intruder test: would a reasonably competent, motivated person with access to resources such as the internet, public libraries, and public records be able to identify an individual from the pseudonymized data?
Assessment Factors:
| Factor | Low Risk (1) | Medium Risk (3) | High Risk (5) |
|---|---|---|---|
| Population uniqueness | Common attributes, large equivalence classes | Moderate uniqueness, some rare combinations | Highly unique attribute combinations |
| Auxiliary information availability | No public datasets linkable | Some public records overlap | Rich public profiles (social media, registers) |
| Identifier entropy | High entropy (e.g., UUID) | Medium entropy (e.g., hashed email) | Low entropy (e.g., hashed phone number) |
| Dataset richness | Few quasi-identifiers | Moderate quasi-identifiers (5-10) | Many quasi-identifiers (>10) |
| Temporal precision | Yearly or coarser | Monthly | Daily or finer |
| Geographic precision | Country level | Region/city level | Postcode or finer |
| Domain sensitivity | Low sensitivity domain | Moderate sensitivity | Health, financial, political data |
Quantitative Risk Models
Prosecutor Model
The attacker knows a specific individual is in the dataset and attempts to find their record. Risk = 1 / k, where k is the size of the equivalence class containing the target record.
Journalist Model
The attacker does not know if the target is in the dataset but attempts to re-identify any individual for a story. Risk = max(1/k) across all equivalence classes (the smallest group is most vulnerable).
Marketer Model
The attacker attempts to re-identify as many individuals as possible. Risk = (1/n) * Σ(1/k_i) — the average re-identification probability across all records.
Linkage Attack Categories
| Attack Type | Description | Mitigation |
|---|---|---|
| Record linkage | Matching pseudonymized records with identified records in external datasets using quasi-identifiers | Generalize quasi-identifiers, enforce k-anonymity |
| Attribute linkage | Inferring sensitive attributes from group characteristics even without identifying the individual | Enforce l-diversity on sensitive attributes |
| Table linkage | Determining that an individual is present in a sensitive dataset (membership inference) | Apply differential privacy, enforce t-closeness |
| Composition attack | Combining multiple independent pseudonymized releases to narrow equivalence classes | Track cumulative privacy budget, limit releases |
| Temporal linkage | Linking records across time periods using behavioral patterns or trajectory data | Apply temporal generalization, add noise to timestamps |
Risk Scoring Methodology
Step 1: Technique Assessment (0-100)
Score the pseudonymization technique on five properties:
| Property | Weight | Score Range |
|---|---|---|
| Key/mapping security | 30% | 0-100 based on key management maturity |
| Brute-force resistance | 25% | 0-100 based on input entropy and computational cost |
| Cross-dataset unlinkability | 20% | 0-100 based on determinism and key reuse |
| Implementation maturity | 15% | 0-100 based on library validation and audit history |
| Operational resilience | 10% | 0-100 based on key rotation, backup, disaster recovery |
Step 2: Data Environment Assessment (0-100)
Score the data environment on re-identification risk factors:
| Factor | Weight | Score Range |
|---|---|---|
| Population uniqueness | 25% | 0-100 (higher = more unique = higher risk) |
| Auxiliary data availability | 25% | 0-100 (higher = more auxiliary data = higher risk) |
| Quasi-identifier count | 20% | 0-100 based on number and granularity |
| Dataset size | 15% | 0-100 (smaller datasets = higher risk per record) |
| Release frequency | 15% | 0-100 (more frequent releases = higher composition risk) |
Step 3: Residual Risk Score
Residual Risk = Data Environment Risk × (1 - Technique Score / 100)
| Residual Risk Range | Classification | Action Required |
|---|---|---|
| 0-15 | Low | Acceptable. Document and monitor |
| 16-35 | Moderate | Additional controls recommended |
| 36-60 | High | Technique upgrade or data reduction required |
| 61-100 | Very High | Processing should not proceed without anonymization |
Prism Data Systems AG — Assessment Summary
| Data Flow | Technique | Technique Score | Environment Risk | Residual Risk | Classification |
|---|---|---|---|---|---|
| Analytics pipeline | HMAC-SHA256 (keyed) | 78 | 42 | 9.2 | Low |
| Support dashboard | FPE (FF3-1) | 82 | 35 | 6.3 | Low |
| Payment processing | Tokenization (vault) | 91 | 55 | 5.0 | Low |
| Test environments | Synthetic replacement | 95 | 20 | 1.0 | Low |
| Research exports | Counter + k-anonymity | 72 | 58 | 16.2 | Moderate |
| ML training data | HMAC + differential privacy | 85 | 48 | 7.2 | Low |
The research exports flow received a "Moderate" classification due to the combination of rich quasi-identifiers in the exported dataset and the counter-based technique's susceptibility to record linkage. Remediation: apply l-diversity (l≥3) on sensitive attributes and generalize quasi-identifiers to achieve k≥11 before export.
Key Regulatory References
- GDPR Article 4(5) — Definition of pseudonymization
- GDPR Recital 26 — Anonymization vs. pseudonymization distinction
- GDPR Article 25(1) — Pseudonymization as a data protection by design measure
- GDPR Article 32(1)(a) — Pseudonymization as a security measure
- GDPR Article 89(1) — Pseudonymization for research processing safeguards
- ENISA "Pseudonymisation Techniques and Best Practices" (November 2019)
- ICO Anonymisation Code of Practice — Motivated intruder test
- Article 29 Working Party Opinion 05/2014 (WP216) on Anonymisation Techniques
- EDPB Guidelines 4/2019 on Article 25 — Pseudonymization as by-design measure
Selecting Privacy-Enhancing Technologies
Overview
Privacy-Enhancing Technologies (PETs) are technical measures that protect personal data during collection, processing, storage, and sharing. The GDPR does not prescribe specific technologies but requires "appropriate technical measures" (Article 25(1), Article 32(1)) to implement data protection principles. The European Commission's February 2025 communication on PETs and the ENISA 2023 report on engineering privacy by design with PETs provide regulatory context for PET adoption.
Selecting the right PET depends on the processing scenario, data sensitivity, computational requirements, and the specific privacy property needed (confidentiality, anonymity, unlinkability, or transparency).
PET Taxonomy
1. Differential Privacy (DP)
Privacy property: Statistical indistinguishability — the output of an analysis is approximately the same whether or not any individual's data is included.
| Characteristic | Detail |
|---|---|
| Approach | Add calibrated noise to query results or model gradients |
| Privacy guarantee | Mathematically provable (epsilon, delta) bounds |
| Data utility | Configurable via epsilon; lower epsilon = more privacy, less utility |
| Performance overhead | Minimal for query-time noise; moderate for DP-SGD training |
| Maturity | Production-ready (Apple, Google, US Census Bureau) |
| GDPR relevance | Supports Recital 26 anonymization when epsilon is sufficiently small; Art. 25(1) by design measure |
Best for: Statistical analytics, aggregate reporting, ML model training on sensitive data.
Libraries: Google DP Library, OpenDP, IBM diffprivlib, PyDP.
2. Homomorphic Encryption (HE)
Privacy property: Computation on encrypted data — results are decrypted only by the data owner.
| Characteristic | Detail |
|---|---|
| Approach | Encrypt data such that mathematical operations on ciphertext produce valid results when decrypted |
| Privacy guarantee | Cryptographic (lattice-based hardness assumptions) |
| Data utility | Exact results (no noise added) |
| Performance overhead | High: 1,000x-1,000,000x slower than plaintext, depending on scheme and operation |
| Maturity | Emerging for production; active research on performance optimization |
| GDPR relevance | Art. 32(1)(a) encryption as security measure; enables processing without exposing plaintext to processor |
Schemes:
- BFV (Brakerski/Fan-Vercauteren): Integer arithmetic, exact. Suitable for counting, matching.
- BGV (Brakerski-Gentry-Vaikuntanathan): Integer arithmetic with modulus switching. Efficient for deep circuits.
- CKKS (Cheon-Kim-Kim-Song): Approximate arithmetic on real/complex numbers. Suitable for ML inference.
- TFHE (Torus FHE): Fast bootstrapping, gate-by-gate evaluation. Suitable for boolean circuits.
Libraries: Microsoft SEAL, IBM HELib, Google FHE (Fully Homomorphic Encryption transpiler), TFHE-rs, OpenFHE.
3. Secure Multi-Party Computation (SMPC)
Privacy property: Joint computation without revealing individual inputs — each party learns only the final result.
| Characteristic | Detail |
|---|---|
| Approach | Secret sharing, garbled circuits, or oblivious transfer protocols |
| Privacy guarantee | Information-theoretic (for secret sharing) or computational (for garbled circuits) |
| Data utility | Exact results |
| Performance overhead | Moderate to high; depends on circuit complexity and number of parties |
| Maturity | Production deployments in financial services, healthcare consortia |
| GDPR relevance | Enables joint controllership (Art. 26) analytics without data sharing; supports data minimization (Art. 5(1)(c)) |
Frameworks: MP-SPDZ, CrypTen (Meta), MOTION, ABY/ABY3, Sharemind.
4. Federated Learning (FL)
Privacy property: Model training without centralizing data — raw data never leaves the data owner's environment.
| Characteristic | Detail |
|---|---|
| Approach | Distribute model training to data owners; aggregate only model updates (gradients) |
| Privacy guarantee | Data locality (raw data stays on device); strengthened with secure aggregation and DP |
| Data utility | Depends on data distribution across participants (IID vs non-IID) |
| Performance overhead | Communication overhead for gradient exchange; multiple training rounds |
| Maturity | Production at Google (Gboard), Apple (Siri), hospitals (NVIDIA FLARE) |
| GDPR relevance | Supports Art. 5(1)(c) minimization; reduces cross-border transfer requirements (Chapter V) |
Frameworks: TensorFlow Federated, PySyft (OpenMined), NVIDIA FLARE, Flower, FedML.
5. Zero-Knowledge Proofs (ZKP)
Privacy property: Prove a statement is true without revealing the underlying data.
| Characteristic | Detail |
|---|---|
| Approach | Prover demonstrates knowledge of a secret (e.g., age ≥ 18) without revealing the secret (date of birth) |
| Privacy guarantee | Soundness (false statements cannot be proven) and zero-knowledge (verifier learns nothing beyond the statement's truth) |
| Data utility | Binary verification (proof valid/invalid); no data disclosed |
| Performance overhead | Proof generation: moderate to high; proof verification: fast |
| Maturity | Production in blockchain identity; emerging in enterprise identity verification |
| GDPR relevance | Ultimate data minimization — prove compliance without disclosing data; supports Art. 5(1)(c), Art. 25 |
Systems: zk-SNARKs (Groth16, PLONK), zk-STARKs, Bulletproofs.
6. Trusted Execution Environments (TEE)
Privacy property: Isolated computation — data is processed inside a hardware-protected enclave that even the system administrator cannot access.
| Characteristic | Detail |
|---|---|
| Approach | Hardware-enforced isolation using secure enclaves (Intel SGX, AMD SEV, ARM TrustZone, AWS Nitro) |
| Privacy guarantee | Hardware-based attestation and memory encryption |
| Data utility | Full computation capability inside enclave |
| Performance overhead | Low to moderate; limited enclave memory may require data streaming |
| Maturity | Production: Azure Confidential Computing, AWS Nitro Enclaves, Google Confidential VMs |
| GDPR relevance | Art. 32(1)(a) encryption in processing; Art. 28 processor guarantees; protects against insider threats |
Use-Case Matching Matrix
| Use Case | DP | HE | SMPC | FL | ZKP | TEE | Recommended Primary |
|---|---|---|---|---|---|---|---|
| Aggregate analytics on customer data | 5 | 2 | 3 | 2 | 1 | 3 | Differential Privacy |
| ML model training on distributed hospital data | 3 | 1 | 2 | 5 | 1 | 3 | Federated Learning |
| Credit scoring without sharing financial records | 2 | 3 | 5 | 2 | 2 | 3 | Secure MPC |
| Age verification without revealing date of birth | 1 | 1 | 1 | 1 | 5 | 2 | Zero-Knowledge Proofs |
| Processing encrypted customer data in third-party cloud | 2 | 5 | 2 | 1 | 1 | 4 | Homomorphic Encryption |
| Anti-money laundering across banks | 2 | 2 | 5 | 3 | 2 | 3 | Secure MPC |
| Private genomic analysis | 3 | 4 | 3 | 3 | 1 | 4 | HE + TEE combination |
| Confidential inference on cloud-hosted ML models | 2 | 4 | 2 | 1 | 1 | 5 | Trusted Execution Environment |
| Privacy-preserving surveys and polls | 5 | 1 | 3 | 1 | 2 | 1 | Differential Privacy |
| Secure keyword search on encrypted database | 1 | 4 | 1 | 1 | 1 | 4 | Homomorphic Encryption |
Scores: 1 = poor fit, 5 = excellent fit
PET Selection Decision Framework
Step 1: Define Privacy Requirements
- What data protection principle is primary? (minimization, confidentiality, anonymization, unlinkability)
- Is the goal to protect data in transit, at rest, or during processing?
- Is the adversary model honest-but-curious or malicious?
- Are there cross-organizational boundaries (multiple controllers/processors)?
Step 2: Assess Constraints
- What is the acceptable performance overhead? (real-time vs batch)
- What is the computational budget? (HE requires significantly more resources)
- How many parties are involved? (SMPC scales with party count)
- Is there a hardware trust anchor available? (TEE requires specific hardware)
- What is the team's cryptographic engineering expertise?
Step 3: Evaluate Combinations
Many production deployments combine PETs for defense in depth:
| Combination | Benefit |
|---|---|
| FL + DP | Federated learning with differentially private gradient updates prevents gradient inversion attacks |
| FL + Secure Aggregation | Aggregation server never sees individual gradients |
| HE + TEE | Process encrypted data inside an enclave for double protection |
| SMPC + DP | Add noise to SMPC outputs for formal anonymization guarantee |
| ZKP + TEE | Prove computation was performed correctly inside an enclave without revealing inputs |
Step 4: GDPR Alignment Verification
- Does the selected PET satisfy Article 25(1) data protection by design?
- Does it implement a specific Article 32(1) security measure?
- Does it support the data minimization principle under Article 5(1)(c)?
- If anonymization is claimed, does it meet Recital 26 standards against singling out, linkability, and inference?
- Is the PET's privacy guarantee documented sufficiently for DPIA (Article 35)?
Key Regulatory References
- GDPR Article 25(1) — Data protection by design
- GDPR Article 32(1)(a) — Encryption and pseudonymisation as security measures
- GDPR Article 5(1)(c) — Data minimization principle
- GDPR Recital 26 — Anonymous information
- GDPR Recital 78 — Technical measures for data protection
- ENISA Report: Data Protection Engineering (2022)
- European Commission Communication on PETs (February 2025)
- EDPB Guidelines 4/2019 on Article 25 Data Protection by Design and by Default
Sponsors
Details
- Marketplace
- mukul975/Privacy-Data-Protection-Skills
- Source
- plugins/privacy-by-design-skills
- Author
- Mahipal
- Category
- Engineering
- License
- Apache-2.0
- Keywords
- privacy-by-design, pets, anonymization, federated-learning, homomorphic-encryption
More from mukul975/Privacy-Data-Protection-Skills
All 21- privacy-engineering-skillsv1.0.0
- privacy-impact-assessment-skillsv1.0.0
- privacy-skills-completev1.0.0
- records-of-processing-skillsv1.0.0
- us-state-privacy-skillsv1.0.0
- vendor-privacy-management-skillsv1.0.0
- ai-privacy-governance-skillsv1.0.0
- children-privacy-skillsv1.0.0
- consent-management-skillsv1.0.0
- cookie-consent-skillsv1.0.0
- cross-border-transfers-skillsv1.0.0
- data-breach-response-skillsv1.0.0
- data-classification-skillsv1.0.0
- data-retention-skillsv1.0.0
- data-subject-rights-skillsv1.0.0
- employee-privacy-skillsv1.0.0
- gdpr-compliance-skillsv1.0.0
- global-privacy-regulations-skillsv1.0.0
- healthcare-privacy-skillsv1.0.0
- privacy-audit-skillsv1.0.0