LLMs Can Now Deanonymize Online Users with 90% Precision
Researchers demonstrate that large language models can effectively deanonymize online users by analyzing their writing style and content across platforms, achieving 68% recall at 90% precision.
Why it matters
This research represents a significant escalation in online deanonymization capabilities, with practical implications for both legitimate investigations and potential misuse.
Key Points
- 1LLM-based pipeline for linking anonymous user accounts across platforms
- 2Extracts identity-revealing patterns from unstructured text posts using LLMs
- 3Achieves 68% recall at 90% precision, significantly outperforming traditional methods
Details
The research team developed a three-stage LLM-based system for online deanonymization. It uses LLMs to extract personal clues, writing style patterns, and other identity-revealing information from user posts. The system then searches through candidate matches across platforms and performs detailed comparison reasoning to determine if they represent the same person. The approach was tested on real-world scenarios like linking Hacker News users to their LinkedIn profiles, achieving much higher accuracy than traditional methods which 'stay near 0%'. This demonstrates that pseudonyms and anonymous usernames have become less effective as privacy protection, as LLMs can now automate the process of linking accounts at scale using only public writing samples.
No comments yet
Be the first to comment