Researchers demonstrate that large language models (LLMs) can effectively deanonymize users across online platforms by analyzing pseudonymous text data, achieving high precision and recall. The attack pipeline extracts identity features, searches via embeddings, and verifies matches, outperforming traditional methods that rely on structured data. This raises significant privacy concerns for online anonymity.
Background
Deanonymization attacks traditionally rely on structured data like in the Netflix Prize, but LLMs enable new approaches using unstructured text from online profiles and conversations. This research highlights evolving threats to user privacy in digital spaces.
- Source
- Lobsters
- Published
- Mar 26, 2026 at 02:06 PM
- Score
- 9.0 / 10