Pseudonymization

Pseudonymization targets direct identifiers—fields that uniquely pinpoint an individual without additional information. These include:

• Name, email address, national ID number
• Phone number, passport number, biometric templates
• IP address, device fingerprints, vehicle registration

Indirect identifiers (quasi-identifiers) like age, ZIP code, or occupation are not replaced but must be assessed for singling-out risk when combined. The GDPR explicitly distinguishes pseudonymized data from anonymous data based on the existence of the mapping key.

Two dominant pseudonymization methods serve different use cases:

Tokenization replaces identifiers with randomly generated tokens, storing the mapping in a secure, isolated vault. The token has no mathematical relationship to the original value, making reversal impossible without vault access.

Cryptographic hashing (SHA-256, bcrypt) applies a one-way function. However, unsalted hashes are vulnerable to rainbow table attacks. Adding a secret salt stored separately from the data satisfies pseudonymization requirements by ensuring reversibility is only possible with the supplementary secret.

The defining characteristic of pseudonymization under GDPR is the technical and organizational separation of the pseudonymized dataset from the re-identification key. Key principles:

• The mapping table must reside in a separate access-controlled system
• No single administrator should have access to both the data and the key
• Attribute-Based Access Control (ABAC) policies enforce that data scientists querying pseudonymized records cannot access the re-identification vault
• This separation transforms the data from 'personal data' to 'pseudonymized data' for the processing party without the key

Pseudonymization does not eliminate re-identification risk—it mitigates it. Residual risks include:

• Linkage attacks: Combining pseudonymized records with external datasets using quasi-identifiers (e.g., Sweeney's 87% U.S. population re-identification using ZIP, gender, DOB)
• Inference attacks: Deriving identity from behavioral patterns or rare attribute combinations
• Key compromise: Unauthorized access to the mapping vault

Effective pseudonymization requires ongoing re-identification risk assessments and monitoring of auxiliary data availability.

In machine learning workflows, pseudonymization enables privacy-preserving model training while maintaining data utility:

• Training data is pseudonymized before ingestion into feature stores
• Differential privacy can be layered on top to bound information leakage from model outputs
• Pseudonymized data that is not reasonably likely to be re-identified by the controller may benefit from exemptions under GDPR Art. 6(4) for compatible further processing
• Combined with federated learning, pseudonymization ensures raw identifiers never leave edge devices

A critical legal and technical distinction:

Pseudonymization: Reversible with additional information. Data remains 'personal data' under GDPR but with reduced risk. The controller holding the key can re-identify.

Anonymization: Irreversible. The data can no longer be attributed to a specific data subject by any means reasonably likely to be used. Truly anonymized data falls outside GDPR scope entirely.

Key test: If the controller can re-identify individuals using 'all means reasonably likely,' the data is pseudonymized, not anonymized. The EU's Article 29 Working Party applies a strict 'means reasonably likely' standard including technological developments and motivated intruder scenarios.

The irreversible process of transforming personal data so that the data subject is no longer identifiable. Unlike pseudonymization, the link to the original identity is permanently destroyed, and the data falls outside the scope of GDPR. Techniques include generalization, suppression, and randomization. True anonymization is difficult to achieve in the age of big data due to linkage attacks.

The process of substituting a sensitive data element with a non-sensitive equivalent (token) that has no extrinsic meaning. The mapping between the original data and the token is stored in a secure, centralized token vault. Unlike pseudonyms generated by hashing, tokens are typically random and preserve format, making them ideal for protecting payment card data and PII in non-production AI environments.

A technique that creates a structurally similar but inauthentic version of data by obscuring specific fields. Common methods include:

• Static Data Masking (SDM): Masks data at rest in non-production databases.
• Dynamic Data Masking (DDM): Masks data in real-time based on user privileges.
• Deterministic Masking: Replaces a value with the same masked value consistently across systems, preserving referential integrity.

A 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). Achieved through generalization (replacing specific values with ranges) and suppression (removing outliers). Vulnerable to homogeneity attacks when sensitive attributes within a k-group are identical, and background knowledge attacks.

A mathematical framework that injects calibrated statistical noise into query results or model training. It provides a formal guarantee: the presence or absence of any single individual's data is indistinguishable in the output. Governed by the privacy loss parameter epsilon (ε)—lower values mean stronger privacy. Implemented via the Laplace mechanism for numerical queries and the exponential mechanism for categorical outputs.

A cryptographic scheme enabling computation directly on ciphertext. The result, when decrypted, matches the output of operations performed on the plaintext. Partially Homomorphic Encryption (PHE) supports only addition or multiplication. Fully Homomorphic Encryption (FHE) supports arbitrary computation but remains computationally intensive, making it suitable for privacy-preserving inference on sensitive data.