Anonymization

A property ensuring each released record is indistinguishable from at least k-1 other records based on quasi-identifiers (e.g., ZIP code, age, gender).

• Mechanism: Generalization (replacing specific values with ranges) and suppression (redacting outliers).
• Weakness: Vulnerable to homogeneity attacks (all k records share the same sensitive value) and background knowledge attacks.
• Example: A dataset where every combination of ZIP code, age, and gender appears at least 5 times achieves k=5 anonymity.

An extension of k-anonymity that requires l distinct sensitive values within each equivalence class, mitigating homogeneity attacks.

• Distinct l-diversity: Ensures at least l different sensitive attributes per group.
• Entropy l-diversity: Requires the entropy of sensitive values to exceed a threshold.
• Limitation: Does not protect against skewness attacks where one sensitive value dominates the distribution.

Refines l-diversity by requiring the distribution of a sensitive attribute in any equivalence class to be within a threshold t of its global distribution.

• Uses Earth Mover's Distance to measure the difference between distributions.
• Prevents skewness attacks and similarity attacks that l-diversity misses.
• Trade-off: Stricter t values significantly reduce data utility for analysis.

A mathematical framework that injects calibrated noise into query results, guaranteeing that the presence or absence of any single individual's record is statistically indistinguishable.

• Quantified by epsilon (ε) — the privacy loss parameter. Lower ε = stronger privacy.
• Laplace Mechanism: Adds noise drawn from a Laplace distribution scaled to sensitivity/ε.
• Composition: Privacy loss accumulates across multiple queries, requiring careful budget tracking.

Replaces sensitive identifiers with surrogate values (tokens) that retain format and referential integrity without exposing original data.

• Format-preserving encryption: Ciphertext matches the format of the original plaintext.
• Vault-based tokenization: A secure lookup table maps tokens back to original values, stored separately.
• Irreversible masking: One-way hashing or character substitution with no recovery path — true anonymization if salts are destroyed.

Replaces direct identifiers with pseudonyms — artificial identifiers that break the obvious link to a data subject but remain reversible with additional information.

• Critical distinction: Pseudonymization is not anonymization under GDPR. The data remains personal if re-identification is possible.
• Common techniques: Hashing with secret salt, encryption with key management, lookup tables.
• Governance requirement: The additional information (key, mapping table) must be kept separately and subject to technical controls.

A mathematical framework that injects calibrated statistical noise into query results or training data. It guarantees that the presence or absence of any single individual's record is indistinguishable, quantified by the privacy loss parameter epsilon (ε). Unlike simple anonymization, differential privacy provides a formal, quantifiable privacy guarantee that holds even against adversaries with arbitrary auxiliary information.

A data privacy property ensuring that each released record is indistinguishable from at least k-1 other records with respect to quasi-identifiers (e.g., ZIP code, age, gender). This limits the granularity of information to prevent singling out individuals. However, k-anonymity is vulnerable to homogeneity attacks and background knowledge attacks, making it a foundational but insufficient standalone anonymization technique.

The statistical probability that an attacker can link anonymized records back to a specific real-world individual by cross-referencing quasi-identifiers with external datasets. Key attack vectors include:

• Linkage attacks: Joining anonymized data with public voter rolls or census data
• Prosecutor risk: The probability of re-identifying a specific known individual
• Journalist risk: The probability that any record in the dataset can be re-identified

A privacy attack where an adversary determines whether a specific individual's record was included in the training dataset of a machine learning model. Attackers exploit differences in model confidence between seen and unseen data. This is a critical failure mode for anonymization: even if direct identifiers are removed, the model's behavior can leak membership information through overfitting signals.

The algorithmic creation of artificial datasets that retain the statistical properties, correlations, and structure of real-world data without containing actual individual records. When properly validated, synthetic data achieves the legal threshold of anonymization by severing the direct link to real individuals. Common generative architectures include GANs, VAEs, and diffusion models.

A core privacy principle under Article 5(1)(c) of GDPR mandating that data collection and processing be limited to what is strictly necessary for a specific purpose. Anonymization is the ultimate expression of data minimization—once data is truly anonymized, it falls outside the scope of data protection regulations entirely, freeing it for secondary analysis and model training.