IAPP AIGP · Body of Knowledge · Domain II.B

De-Identification & Privacy-Enhancing Technologies

Anonymisation, pseudonymisation, differential privacy and the full PET toolbox — what each technique is, how they compare, and the exam traps that separate them.

BoK Domain II.B · Data Governance 8 comparison tables 20 flashcards 8-question self-test GDPR · PDPA · ISO lens
10
De-identification techniques
6
PET categories
4
Re-identification attacks
3
k-anon / l-diversity / t-closeness
1
Golden rule: pseudonymised ≠ anonymised
Orientation

0 The De-Identification Landscape at a Glance

Every technique in this module answers one question: how do we reduce the link between data and an identifiable individual, and what obligations remain after we do? The spectrum runs from fully identified data at one end to (ideally) irreversible anonymisation at the other — with pseudonymisation and the privacy-enhancing technologies in between.[AIGP BoK II.B]

De-Identification & PETs AIGP BoK Domain II.B 1 · The Spectrum identified → pseudonymised → anonymised the golden line 2 · Anonymisation aggregation · generalisation noise · swapping · hashing k-anonymity family 3 · Pseudonymisation tokenisation · encryption EDPB 01/2025 STILL personal data 4 · The PET Toolbox differential privacy · FL synthetic data · SMPC · HE secure enclaves · ZKP 5 · Re-Identification linkage · singling-out inference · unmasking motivated intruder test 6 · Law & Practice GDPR R26 · PDPC Guide ISO 27559 · PDPA AI guide EU AI Act Art. 10(5) exam spine: which technique · what remains in scope · what re-identification risk survives
BoK Domain II.B — The de-identification spectrum

1 The De-Identification Spectrum & the Golden Line

De-identification is not one thing. It is a spectrum of techniques with different strengths, different re-identification risks, and — critically — different legal status. The exam tests whether you know exactly where each technique sits and what obligations survive it.[AIGP BoK II.B]

1.1 Reading the spectrum

PositionWhat it meansLegal statusExample
Identified dataDirect identifiers present (name, NRIC, email, phone)Personal data — full regime appliesCustomer CRM records
De-identified (generic term)Direct identifiers removed or altered by any techniqueDepends on the technique and residual risk — not automatically out of scopeAny dataset after an ID-stripping pass
PseudonymisedIdentifiers replaced with keys; re-linkage possible with the keyStill personal data — GDPR Art. 4(5); PDPA applies likewise[GDPR; EDPB 01/2025]Research database keyed by patient ID held separately
AnonymisedIdentifiers removed and re-identification reasonably prevented, irreversiblyOut of scope of GDPR/PDPA if genuinely achieved[GDPR R26; PDPC]Aggregated statistics with no residual link
The golden line (memorise this first)Anonymisation is the only position on the spectrum that exits the data-protection regime. Pseudonymisation is a safeguard, not an exit. Any exam answer claiming pseudonymised data is no longer personal data is wrong — in both the EU and Singapore contexts.[GDPR; EDPB 01/2025]

1.2 Why "anonymised" is a claim, not a label

The EDPB's Article 29 Working Party Opinion 05/2014 (still the standard reference) treats anonymisation as a result to be tested against three criteria, not a box to tick: singling out (can one individual be isolated in the dataset?), linkage (can two records about the same individual be connected?), and inference (can a value about an individual be deduced with significant probability?). If any criterion is reasonably achievable, the data is not truly anonymised.[WP29 05/2014]

S

Singling out

Isolating one individual's record from the crowd — e.g., a row that is unique in the dataset.

L

Linkage

Connecting two or more records that belong to the same individual across tables or datasets.

I

Inference

Deducing an individual's attribute with significant probability from the remaining data.

Exam anchorThe AIGP exam regularly asks which test establishes whether anonymisation has been achieved. Answer: the motivated intruder test — whether a reasonably competent outsider, with access to public resources, could re-identify individuals. This is the standard the EDPB applies to all three criteria above.[WP29 05/2014]

1.3 The Singapore framing

Singapore's PDPC Guide to Basic Anonymisation (first published 2018, updated 2024) walks organisations through a 5-step anonymisation process: (1) identify direct and indirect identifiers, (2) apply de-identification techniques, (3) assess re-identification risk, (4) determine the residual risk of re-identification, and (5) decide on the appropriate data-sharing safeguards. PDPC also publishes a free Data Anonymisation Tool based on the guide.[PDPC 2024]

The PDPC's Advisory Guidelines on the PDPA for Selected Topics (revised 2024) clarify that data protection obligations continue to apply to anonymised data if re-identification remains reasonably possible — in practice, organisations are expected to assess residual risk rather than rely on the label "anonymised".[PDPC AG 2024]

The word "anonymised" is a legal conclusion, not a technical description of what you did to the data. Two organisations can apply the same masking tool; one has genuinely anonymised (no residual risk) and the other has not (key stored, low-but-real linkage risk). The exam tests this distinction relentlessly.[EDPB 01/2025; PDPC]
BoK Domain II.B — Techniques that (aim to) remove the link

2 Anonymisation Techniques & the k-Anonymity Family

Anonymisation techniques remove or distort the link between data and individuals. Each one trades utility against residual risk — and none is a silver bullet. The exam wants you to recognise the technique from a description and know its weakness.[WP29 05/2014; ISO 20889]

2.1 The core techniques — comparison table

TechniqueWhat it doesTypical useKey weakness
AggregationCombines records into groups and reports only group statistics (totals, averages)Official statistics; research disclosuresSmall group sizes still leak individuals; loses record-level detail[WP29; ISO 20889]
Generalisation (incl. banding / binning)Broadens attribute values: exact age → age band; postcode → district; job title → industry sectork-anonymity datasets; public health dataReduces data precision; heavy generalisation can destroy analytical value[Sweeney 2002; PDPC]
Noise additionAdds random perturbation to values so individual records are distorted but distributions surviveStatistical databases; location dataDeterministic attributes (age, gender) unaffected; can bias downstream analysis
Data swapping / permutationExchanges attribute values between records to break the record-to-person linkCensus-style disclosure controlOriginal values may still appear in combination elsewhere; ordering attacks
Masking / redactionRemoves or blurs parts of values (e.g., NRIC S****123A; names → initials)Screen redaction; test datasetsPartial identifiers often remain linkable — masking alone rarely anonymises
Hashing / keyed hashingReplaces identifiers with hash digests; keyed hashing adds a secret so values can't be brute-forcedJoining datasets without exposing raw IDsUnkeyed hashes of low-entropy values (names, NRICs) are trivially reversed[WP29]
k-anonymity (model)Ensures every record is indistinguishable from at least k−1 others on quasi-identifiersPublished microdataFails on homogeneity and background-knowledge attacks (see l-diversity, t-closeness)

2.2 The k-anonymity family — the exam's favourite microdata models

k-anonymity (Sweeney, 2002): a dataset satisfies k-anonymity if every combination of quasi-identifiers (attributes like age, gender, postcode that can link back to a person) appears in at least k records. The classic failure: homogeneity attacks — if all k records share the same sensitive value (e.g., all 5 people in a group have the same disease), knowing someone is in the group reveals the disease.[Sweeney 2002]

l-diversity (Machanavajjhala et al., 2007): extends k-anonymity by requiring at least l "well-represented" distinct values for each sensitive attribute within each group — so a group can't be homogeneous on the sensitive value.[l-diversity 2007]

t-closeness (Li et al., 2007): goes further — the distribution of sensitive values in each group must be close (within distance t) to the distribution in the whole dataset. This blocks skewness attacks, which l-diversity misses when the global distribution is itself skewed.[t-closeness 2007]

ModelGuaranteeAttack it blocksLimitation
k-anonymityEach quasi-identifier group has ≥ k recordsRecord linkage / singling outHomogeneity attack — group can share one sensitive value[Sweeney 2002]
l-diversityEach group has ≥ l distinct sensitive valuesHomogeneity attackSkewness attack — skewed global distribution defeats "distinctness"[l-diversity 2007]
t-closenessGroup distribution of sensitive values ≈ global distribution (within t)Skewness attackUtility loss; hard to tune t; not yet a formal privacy guarantee
Exam memory hookk-anonymity = "k of us look the same". l-diversity = "l different secrets in the group". t-closeness = "the group's secret mix ≈ the whole dataset's mix". Each generation fixes the previous one's blind spot.[k-anon → t-close]
The canonical re-identification case the AIGP materials reference: in 1997, Massachusetts Governor William Weld's medical records were re-identified from supposedly anonymised group-insurance data by linking birth date, ZIP code and sex (quasi-identifiers) against the voter register — the very attack k-anonymity was invented to prevent (Sweeney's own motivating example).Source: Sweeney, "k-Anonymity: A Model for Protecting Privacy" (2002); widely cited in IAPP materials
Deep-dive: k-anonymity worked example (exam-ready)

The setup

A hospital wants to publish a research table with columns Age band, District, Disease. Before k-anonymity, a row "34, Clementi, HIV+" with no other 34-year-old Clementi resident in the table is immediately linkable to exactly one person via the voter register.

The fix

  • Generalise Age 34 → 30–39 and District Clementi → West (banding + generalisation).
  • Now suppose 5 rows share (30–39, West). The dataset is 5-anonymous: an attacker can narrow you down to 5 people, not 1.

The exam trap

If all 5 rows show HIV+, k-anonymity is satisfied but privacy is destroyed — the homogeneity attack. l-diversity would require ≥2 distinct diseases; t-closeness would require the group's disease mix to resemble the hospital's overall mix.[l-diversity; t-closeness]

Exam anchor

Scenario: "the dataset satisfies k-anonymity but all records in a group share the same sensitive value" → the correct response is l-diversity (or t-closeness for skewness). Recognise the attack from the symptom description.

BoK Domain II.B — The safeguard that keeps obligations

3 Pseudonymisation — What It Is, What It Isn't

Pseudonymisation replaces direct identifiers with artificial identifiers (pseudonyms) so that data can no longer be attributed to a data subject without the use of additional information — which must be kept separately and subject to technical and organisational measures. That last clause is the whole exam: the key exists, so the link survives, so the data stays personal.[GDPR Art. 4(5); EDPB 01/2025]

3.1 Techniques for creating pseudonyms

TechniqueHow it worksReversibilityExam note
TokenisationIdentifier replaced by a random token stored in a secure mapping table (vault)Reversible via the vaultMost common enterprise pattern; the vault is the risk point[EDPB 01/2025]
EncryptionIdentifier encrypted with a key; ciphertext serves as pseudonymReversible with the keyDistinct from tokenisation: deterministic encryption allows matching across datasets — a feature and a risk
Keyed hash (HMAC)Identifier hashed with a secret key; the digest is the pseudonymRecomputable with the keyWithout the key, low-entropy identifiers are brute-forceable; key management is everything[EDPB; WP29]
Random pseudonymsFully random replacement, no mathematical link to the originalOnly via a separately kept mappingStrongest form if the mapping is destroyed → becomes anonymisation
The destroy-the-key escape hatchIf the separately held mapping or key is destroyed irreversibly, pseudonymised data can become anonymised — because the "additional information" that enables attribution no longer exists. This is the exam's favourite edge case: pseudonymisation + key destruction = anonymisation, but only if no other means of re-identification remain.[EDPB 01/2025; GDPR R26]

3.2 What EDPB Guidelines 01/2025 changed

The EDPB adopted Guidelines 01/2025 on Pseudonymisation on 16 January 2025 (public-consultation version). Key exam-relevant positions:[EDPB 01/2025]

Pseudonymisation ≠ anonymisation

Reiterated emphatically: pseudonymised data is personal data for any controller or processor holding the means of attribution, and remains in scope of GDPR (and by analogy PDPA).

It is a risk-reduction measure

Framed as a security and data-minimisation measure (Art. 32 and data-protection-by-design), not an exemption. It supports legitimate-interest balancing and compatible-purpose analysis.

Evidence over labels

Controllers must document the technical and organisational measures protecting the key, the categories of persons with access, and the residual re-identification risk — the audit trail is what regulators will ask for.

Same-data pseudonymisation

Pseudonymising with the same key across datasets still permits linkage by the key-holder — a common mistake in research consortia and a documented exam scenario.

Real-world anchor: NHS Digital's pseudonymisation service for England's health data (introduced in 2023) replaced direct identifiers with pseudonyms and enforced separate, role-based access to the key-holding service. The design was explicitly built around the GDPR Art. 4(5) structure — pseudonyms for analytics, keys quarantined, and every access logged. It is a textbook "safeguard, not escape" implementation.Source: NHS Digital/England policy documentation on the pseudonymisation service (2023); EDPB Guidelines 01/2025 framing

3.3 Pseudonymisation vs anonymisation — the master comparison

PseudonymisationAnonymisation
Reversible?Yes — via separately held key/mappingNo — link destroyed irreversibly
Still personal data?Yes — GDPR Art. 4(5); PDPA applies[GDPR; EDPB]No, if genuinely achieved (high bar)[GDPR R26]
Role in lawSafeguard / security measure (Art. 32); data-minimisation support; lawful-basis enablerExit from the regime; enables free reuse and sharing
Re-identification riskControlled by key security — but the key-holder can always linkResidual risk from linkage/inference attacks; must be tested (motivated intruder)
Typical useAnalytics, research with follow-up, cross-dataset joins, test environmentsOpen-data publication, statistics, long-term archives
BoK Domain II.B — Privacy-Enhancing Technologies

4 The PET Toolbox — Full Comparison

Privacy-Enhancing Technologies (PETs) are the broader family: techniques and tools that minimise personal data use while preserving functionality. They range from the mathematical (differential privacy) to the architectural (federated learning, secure enclaves). The AIGP BoK groups them under data governance in the development lifecycle — expect scenario questions asking "which PET fits this constraint?"[AIGP BoK II.B; PDPC PET Guide]

4.1 The full PET comparison table

PETWhat it doesProtects againstKey tradeoff / limitationFlagship example
Differential privacy (DP)Adds calibrated noise to queries or training so no single record materially changes outputs; governed by privacy budget εMembership inference; singling out; any individual-record disclosurePrivacy–utility tradeoff; ε is a policy choice; noisy small datasets are useless[Dwork 2006; BoK]US Census Bureau 2020 disclosure avoidance
Federated learning (FL)Training moves to the data; only model updates (gradients) leave devices/organisationsCentralisation of raw data; exfiltration of the corpusGradients can leak information (gradient-inversion attacks) — pair with DP and secure aggregation[McMahan 2017; BoK]On-device keyboard prediction (Gboard)
Synthetic dataGenerative models produce statistically similar artificial records; no direct 1:1 subject mappingRe-identification of source subjects (if generation is sound)Can memorise and regurgitate outliers; inherits source bias; utility varies by task[PDPC Synthetic]Healthcare data-sharing pilots (e.g., UK NHS synthetic data initiatives)
Secure multi-party computation (SMPC)Multiple parties compute jointly on their combined data; each learns only the result, never others' inputsDisclosure of inputs during joint computationCommunication/compute overhead; party-collusion analysis; hard to scale[Yao 1982]Cross-bank fraud-network detection without sharing customer data
Homomorphic encryption (HE)Computation performed directly on encrypted data; result decrypts to the correct answerDisclosure to the computing party (even if compromised)High computational overhead; complex to implement correctly[Gentry 2009]Encrypted analytics on cloud-hosted health data
Trusted execution environments (TEE / secure enclaves)Hardware-isolated enclaves where code and data are protected even from the host OSHost compromise; insider access to data in useHardware trust assumptions; side-channel attacks historically[IAPP; TEE]Confidential computing on cloud platforms (SGX/SEV/Nitro)
Zero-knowledge proofs (ZKP)Prove a statement is true (e.g., "age > 18") without revealing the underlying valueOver-disclosure of attributes during verificationComputational cost; narrower use cases (verification, not analytics)Age-verification and credential attestation
Anonymisation / de-identification suiteThe §2 techniques: aggregation, generalisation, noise, swapping, k-anonymity familyDirect identification in published dataUtility loss; residual linkage/inference risk (never zero)[WP29]PDPC Data Anonymisation Tool
Pseudonymisation / tokenisationIdentifier replacement with separately held keys (§3)Casual exposure of identifiers in analyticsKey-holder can always re-link; data remains in scope[GDPR; EDPB]NHS England pseudonymisation service

4.2 PET selection — the exam's decision logic

Q1 — Must the data leave your control?
Architecture first
No → keep raw data local; consider federated learning or secure enclaves so only outputs/updates move.[AIGP BoK II.B]
Q2 — Do multiple parties need to compute jointly?
Collaboration
Yes → SMPC (joint computation, no input sharing) or HE (compute on encrypted data by an untrusted party).
Q3 — Is a formal privacy guarantee required?
Math vs effort
Yes → differential privacy with a documented ε. No formal guarantee needed → de-identification suite (k-anonymity family) may suffice — but re-identification risk must still be assessed.
Q4 — Can the analysis survive artificial data?
Utility check
Yes → synthetic data for testing, augmentation, and sharing. Verify fidelity per task — synthetic ≠ real for every downstream use.[PDPC Synthetic]
Q5 — Is a single attribute the only thing needed?
Minimal disclosure
Yes → zero-knowledge proofs for attribute verification (age, eligibility) without revealing the value.
Combination is normalPETs are complements, not competitors: federated learning + differential privacy + secure aggregation is the canonical stack for cross-institution ML (Apple's and Google's on-device systems). The exam rewards answers that combine PETs rather than picking a single winner.[McMahan 2017; BoK]

4.3 DP vs FL vs synthetic data — the three most-confused PETs

Differential privacyFederated learningSynthetic data
Core mechanismCalibrated noise in outputs/training (formal ε guarantee)Distributed training; only model updates travelGenerate artificial records from learned distributions
What it protectsAny single record's influence on resultsRaw data staying at sourceDirect subject mapping in released datasets
Primary weaknessUtility loss; ε is a value judgment[Dwork 2006]Gradient leakage without DP/secure aggregation[McMahan 2017]Memorisation of outliers; inherited bias[PDPC Synthetic]
Best-fit scenarioStatistical releases and queries over large datasetsModel training across devices/organisationsTesting, augmentation, dataset sharing
Deep-dive: differential privacy in one page

The intuition

A query is differentially private if its output distribution barely changes whether or not any single individual is in the dataset. Add the right amount of random noise and an attacker literally cannot tell if "your" record contributed — the strongest known defence against linkage and inference attacks.

The privacy budget (ε)

ε bounds the information leaked per query. Lower ε = stronger privacy, noisier answers. Budgets accumulate across queries — run too many and the guarantee erodes. ε is set by the organisation; it is a policy decision, not a technical constant. Some systems also track a δ parameter for rare bad events.

Local vs central DP

Central DP: noise added by the trusted curator after collecting data (US Census 2020). Local DP: noise added on-device before anything leaves the user (Apple's keyboard/health analytics) — stronger protection, much noisier.

Exam anchor

"A government wants to publish census statistics such that no individual's data can be inferred, with a tunable privacy guarantee" → differential privacy. "A hospital consortium wants to train a model without moving patient records" → federated learning (with DP as a complement).[Dwork 2006; US Census]

Deep-dive: synthetic data — the 2023 PDPC guide

What the guide says

PDPC's proposed guide (2023) classifies synthetic data as a PET and walks through generation approaches — statistical models, generative models (GANs, diffusion), and agent-based simulation — plus the key governance steps: fidelity validation, re-identification risk assessment of the generator, and disclosure controls on the model itself.

The two failure modes the exam cares about

  • Memorisation: generative models can reproduce near-identical copies of rare or outlier source records — synthetic then is personal data.
  • Bias inheritance: synthetic data inherits every skew of the source data; it does not fix representation problems, it reproduces them.

Exam anchor

"Synthetic data is always privacy-safe" → false. The safety depends on generator design, fidelity controls, and re-identification testing.[PDPC Synthetic 2023]

BoK Domain II.B — Why "de-identified" can fail

5 Re-Identification Attacks & Residual Risk

Every de-identification decision is a bet against re-identification. The exam wants you to recognise the four canonical attack classes and understand that residual risk is measured against a motivated intruder, not a casual browser.[WP29 05/2014; AIGP BoK II.B]

5.1 The four attack classes

Linkage attack

Joining a de-identified dataset with another source (voter register, social media, commercial data) on shared quasi-identifiers to re-identify records.

Example: the 1997 Weld medical-record case — birth date + ZIP + sex joined to the voter list.[Sweeney 2002]

Singling-out attack

Isolating one individual's record from a dataset even without a second source — unique attribute combinations make rows identifiable in isolation.

Countered by k-anonymity (group size ≥ k).[WP29]

Inference attack

Deducing a sensitive attribute with high probability from statistical or aggregate outputs — even when no record is directly exposed.

Membership inference against ML models is the modern form.[WP29; Shokri 2017]

Unmasking / de-pseudonymisation

Recovering identities behind pseudonyms — via the stolen key, brute-forcing weak hashes, or behavioural matching.

Why key management and unkeyed-hash prohibition are governance requirements.[EDPB 01/2025]

5.2 Re-identification risk factors — what raises residual risk

FactorWhy it raises riskMitigation
Rare attribute combinationsUnique rows are singling-out baitGeneralisation; k-anonymity; suppression of small groups
Rich external data environmentsMore public datasets = more linkage fuelMotivated-intruder test against current data landscape
Low-entropy identifiers (names, NRICs, dates)Brute-forceable even when hashed unkeyedKeyed HMAC; tokenisation; never publish unkeyed hashes[WP29]
Longitudinal dataMultiple releases about the same individuals multiply linkage surfacesConsistent release policy; disclosure-control review per release
High-dimension data (location, text)Location traces and free text are near-unique signaturesSpatial generalisation; differential privacy for queries; careful text handling
Key/mapping compromisePseudonymisation collapses instantly if the vault leaksSeparate storage, role-based access, audit logging, encryption at rest[EDPB 01/2025]
The motivated intruder test (define it cold)A standard for assessing residual re-identification risk: would a reasonably competent person, without inside knowledge or specialised equipment, using publicly available resources, be able to re-identify individuals? It is the EDPB's operational test for whether data is genuinely anonymous — and the AIGP exam's definition question in this area.[WP29 05/2014]
The AOL search scandal (2006) is the standard "de-identified but not anonymous" cautionary tale: AOL published 20 million "anonymised" search queries with user IDs. Reporters re-identified user 4417749 — Thelma Arnold of Lilburn, Georgia — purely from search content, including her name typed into her own searches. Query-level data is near-impossible to anonymise; the disclosure was pulled within days.Source: New York Times, "A Face Is Exposed for AOL Searcher No. 4417749" (Aug 2006); widely cited in IAPP training materials
Deep-dive: is hashing anonymisation? (classic exam trap)

No — and here's the reasoning chain

  1. Unkeyed hashes of names/NRICs are trivially reversible by brute force over the small space of possible values (dictionary attack). The "hash" is effectively an encoding, not protection.
  2. Keyed hashing (HMAC) resists brute force, but the key-holder can recompute the original identifier — so the data is pseudonymised, not anonymised.
  3. Even a perfect pseudonym leaves the record linkable across datasets whenever the same key is used (linkage attack).

Exam anchor

"We hashed the NRICs, so the dataset is anonymous" → false: unkeyed hashing is not anonymisation, keyed hashing is pseudonymisation. Either way, obligations remain.[WP29; EDPB]

How the regimes treat the spectrum

6 Regulatory Treatment — GDPR, PDPA, EU AI Act, ISO

De-identification is not just a technical topic — every major regime has a position on it, and the exam tests where each regime draws its lines. Singapore and the EU agree on the core: pseudonymised = in scope, truly anonymised = out.[GDPR; PDPA; EU AI Act]

6.1 The regimes compared

RegimePosition on anonymisationPosition on pseudonymisationKey instrument
EU GDPRRecital 26: principles don't apply to anonymous info; bar is "reasonably likely means" of re-identificationArt. 4(5) definition; Art. 32 security measure; EDPB 01/2025 guidanceGDPR Art. 4(5), Art. 32, Recital 26[GDPR]
Singapore PDPAPDPC Guide to Basic Anonymisation: 5-step process; obligations persist if re-identification reasonably possibleNot a defined term in the Act; treated as a safeguard — data remains personal dataPDPC Guide to Basic Anonymisation (2024); Advisory Guidelines for Selected Topics[PDPC]
EU AI ActArt. 10(5): special-category data for bias detection/monitoring in high-risk systems may be processed "subject to appropriate safeguards for the fundamental rights and freedoms" — including technical measures such as pseudonymisationArt. 10(5) explicitly names pseudonymisation among permitted safeguards; Art. 10(3) requires special-category processing "strictly necessary" and "appropriate safeguards"EU AI Act Art. 10(3), (5)[EU AI Act]
ISO/IEC 27559:2022Framework for de-identification: risk assessment, governance, and lifecycle management of de-identified dataComplements 27559 with the de-identification technique catalogue in ISO/IEC 20889:2018ISO/IEC 27559:2022; ISO/IEC 20889:2018[ISO 27559; ISO 20889]
EU AI Act (data governance for high-risk AI)Anonymisation listed among techniques relevant to training-data governanceEU AI Act Art. 10(2)(b) (training data governance practices incl. de-biasing)
EU AI Act Art. 10(5) — the specific exam citationHigh-risk AI providers may process special-category personal data (sensitive data) solely to ensure bias detection and correction, subject to strict necessity, appropriate safeguards, and technical measures — pseudonymisation is named explicitly as such a measure. This is the AI-Act hook for PETs and a favourite citation question.[EU AI Act Art. 10(5)]

6.2 Why the law cares about the distinction

ConsequencePseudonymised (in scope)Anonymised (out of scope)
Lawful basis needed?Yes — consent, legitimate interests, business improvement, etc.No
Purpose limitation applies?Yes — repurposing requires a new/compatible basisNo — free reuse, sharing, sale
Data-subject rights (access, erasure)?Yes — and the controller must be able to link data to the subject (or justify why not)No — no identifiable subject exists
Breach notification?Yes, if a breach compromises the data/keyGenerally no
DPIA trigger reduction?Pseudonymisation reduces risk and can lower DPIA scope — but doesn't eliminate the needN/A — out of the data-protection analysis entirely
Why regulators insist on this line: if pseudonymised data were treated as anonymous, organisations could dodge every obligation simply by running a tokeniser — while still holding the key that re-identifies everyone. The law follows the capability to re-identify, not the label on the dataset.[EDPB 01/2025]
Easily confused — exam differentiators

7 Compare & Contrast Tables

Five pairs the exam loves to blur. If you can fill these from memory, this domain is yours.

7.1 Anonymisation vs pseudonymisation (legal status)

AnonymisationPseudonymisation
Reversible?No — link destroyedYes — via separately held key
Still personal data?No, if genuinely achievedYes — GDPR Art. 4(5)[GDPR]
Legal effectExits the regimeSafeguard; risk reduction; remains regulated
How to proveMotivated intruder test / WP29 criteriaKey-security documentation and access control[EDPB]

7.2 Differential privacy vs federated learning

Differential privacyFederated learning
Protects againstIndividual-record influence on outputs (formal ε guarantee)Centralisation / movement of raw data
MechanismCalibrated noise in queries or trainingDistributed training; only updates shared
Key weaknessUtility loss; ε is a policy tradeoffGradient leakage; no formal guarantee alone[AIGP BoK II.B]
Flagship exampleUS Census 2020 disclosure avoidanceOn-device keyboard prediction (Gboard)

7.3 k-anonymity vs l-diversity vs t-closeness

k-anonymityl-diversityt-closeness
Requirement≥ k records per quasi-identifier group≥ l distinct sensitive values per groupGroup sensitive distribution within t of global
BlocksSingling out / record linkageHomogeneity attackSkewness attack
Fails againstHomogeneity; background knowledgeSkewness; background knowledgeBackground-knowledge extremes (partial)
CitationSweeney 2002Machanavajjhala 2007Li, Kifer, Machanavajjhala 2007

7.4 SMPC vs homomorphic encryption

SMPCHomomorphic encryption
ModelParties jointly compute; each holds a share of the resultOne party computes over encrypted data with the key held elsewhere
Data staysDistributed among parties (secret-shared)Encrypted at the computing party
Best forJoint analytics without sharing inputsOutsourced computation (cloud) on sensitive data
OverheadCommunication roundsComputation cost

7.5 Personal data vs special-category data (de-identification context)

Personal dataSpecial-category / sensitive data
GDPR regimeArt. 6 lawful basisArt. 9 conditions (plus Art. 6)
ExamplesName, contact, location, IDsHealth, race/ethnicity, biometrics, religion, sexual orientation
AI Act hookGeneral training-data governance (Art. 10(2))Bias detection exception with safeguards incl. pseudonymisation (Art. 10(5))[EU AI Act]
Exam significanceBaseline dutiesHigher bar; pseudonymisation named as an enabling safeguard
Active recall

8 Flashcard Deck — 20 High-Yield Terms

Click a card to flip. Each back face carries a definition plus an exam tip. Shuffle order in your head, not the deck — spaced repetition beats novelty tonight.

Self-test

9 Eight-Question Self-Test

Instant feedback per question; score tallied at the bottom. Best score persists in this browser via localStorage.

Citation ledger

10 Sources Cited