LLMs Reveal Identities Behind Anonymous Accounts
A study by Berlin’s MATS Research and ETH Zurich, published in early March, shows large language models can deanonymize social media accounts at scale. Using semantic embeddings across Hacker News, Reddit, and LinkedIn test accounts, the authors report up to 68% recall at 90% precision, linking anonymous posts to real LinkedIn identities. Researchers warn this poses privacy and free-speech risks though the work used fabricated accounts and requires matching real online profiles.
Key Points
- 1Demonstrates LLMs deanonymize social media accounts with up to 68% recall at 90% precision
- 2Highlights risk to privacy and free speech from scalable identity-linking across public writings
- 3Warns practitioners to limit public textual traces and consider mitigation for semantic-embedding attacks
Scoring Rationale
High research significance and broad applicability, limited by experimental setup using fabricated accounts and real-world validation needs.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Logistics & Shipping problems
