LLMs can unmask pseudonymous users at scale with surprising accuracy
9 hours ago
- #AI
- #deanonymization
- #privacy
- AI can analyze burner accounts to deanonymize users with high success rates (68% recall, 90% precision).
- Pseudonymity, a common privacy measure, is now vulnerable, exposing users to doxxing, stalking, and profiling.
- Researchers used datasets from Hacker News, LinkedIn, Netflix, and Reddit to test deanonymization techniques.
- The findings challenge the assumption that pseudonymity provides adequate privacy protection against targeted deanonymization.
- Techniques included cross-platform references, large language models, and splitting user histories for identification.