How have re-identification attacks evolved from the 2000s to the 2020s?

From simple linkage on tabular data, to inference on rich medical/behavioral data, to AI-powered attacks on spatiotemporal, network, and ML-model data.

The four foundational attack types stayed the same, but the targets grew more complex:

• 2000s — linkage and homogeneity attacks on simple tabular data (e.g. medical records linked to voter rolls).
• 2010s — inference and similarity attacks on complex medical and behavioral datasets.
• 2020s — advanced attacks on spatiotemporal (location), network (social graph), and ML-model data, increasingly driven by AI.

The principles are unchanged; what's escalated is the richness of the data attackers exploit and the power of the tools (machine learning) they bring.

Tip: Defenses designed for 2000s tabular data (k-anonymity) were never built for 2020s behavioral/location/graph data — which is why the field moved to differential privacy and synthetic data.