Data Anonymization.

Data anonymization irreversibly transforms personal data so that individuals can no longer be singled out, linked, or inferred — even by combining the release with outside information. Core techniques are suppression, generalization (e.g. coarsening a birth date to a year), aggregation, perturbation/noise, and randomization, usually evaluated against a formal model: k-anonymity, l-diversity, t-closeness, or differential privacy, which adds calibrated noise under a provable privacy budget (ε).

The bar is high because of a long record of re-identification. Latanya Sweeney famously estimated that 87% of Americans are unique on the combination of 5-digit ZIP, date of birth, and sex, and used public voter rolls to re-identify Massachusetts Governor William Weld's hospital records in "anonymized" state data — the result that motivated k-anonymity. In 2008 Narayanan and Shmatikov de-anonymized the Netflix Prize dataset (ratings of ~500,000 subscribers) by correlating sparse ratings with public IMDb profiles. The 2006 AOL search-log release was traced back to named individuals within days.

Under the GDPR, truly anonymized data falls outside the Regulation (Recital 26), but the Article 29 Working Party's Opinion 05/2014 and regulators such as the EDPB and CNIL require a documented re-identification risk assessment covering singling-out, linkability, and inference using means "reasonably likely to be used." Common pitfalls: relying on hashing or simple pseudonyms, releasing high-dimensional micro-data, or confusing pseudonymization with anonymization.

flowchart TD
  R["Raw personal data"] --> D["Remove direct identifiers"]
  D --> Q["Generalize / suppress quasi-identifiers"]
  Q --> M["Apply privacy model (k-anonymity / differential privacy)"]
  M --> T{"Re-identification risk acceptable?"}
  T -->|"No"| Q
  T -->|"Yes"| P["Publish / share anonymized dataset"]