Overview
This series examines Large-scale online deanonymization with LLMs and the privacy problem it exposes: semantic traces can remain linkable after names and handles are removed, while extraction, search, reasoning, and calibration lower matching costs. It moves from quasi-identifiers and persistence to bounded experimental results, then to release, access, validation, and monitoring decisions. The evidence supports cheaper, more scalable matching in selected settings—not a universal identification rate—and leaves transfer across populations and future defenses unresolved.
Learning outcomes
- Removing names and handles can reduce obvious lookup routes without removing semantic linkability.
- LLM-assisted matching lowers investigation costs through extraction, search, reasoning, and calibrated abstention.
- Reported precision and recall are conditional on datasets, candidate pools, prevalence, thresholds, and verification.
- Privacy protection requires purpose-based transformation, controlled access, validation, monitoring, and accountable revision.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does Large-scale online deanonymization with LLMs show?
It shows that LLM-assisted systems can link selected pseudonymous profiles to identities under tested conditions while lowering investigation costs. It does not establish a universal identification rate.
Why is removing names not enough for pseudonymity?
Removing names is not enough because quasi-identifiers, relationships, repeated details, and stable pseudonyms can preserve linkability through auxiliary data.
How should precision and recall be interpreted in deanonymization experiments?
Precision measures how many returned matches are correct, while recall measures how many true matches are found. Both depend on the dataset, candidate pool, threshold, and prevalence of real matches.
What should organizations do before releasing pseudonymous text?
Organizations should define purpose and recipients, test privacy and utility, choose access controls, and monitor for new linkage opportunities. Validation reduces uncertainty but cannot prove permanent safety.