Behavioural Fingerprinting for AI Anonymous Dialogue Decode
IDShield
In the digital era, LLMs often overlook privacy. The novel tool "UMA" proactively identifies hidden compliance gaps in anonymous AI interactions, enabling companies to audit systems and align with regulations like the GDPR.
“Anonymous” human–LLM chats at risk of unintended privacy exposure
Key Features & Advantages
FIRST SOLUTION: End-to-end automated compliance audit tool
Cross-dialogue re-identification of an individual by behavioural and linguistic fingerprints
Detect over 92% of re-identification risk and improve verification performance
Potentially the standard compliance-check method for AI service providers
This invention presents UMA (Uncertainty-aware Multi-aspect Attribution), a novel AI security framework designed to de-anonymize users across disparate, unauthenticated LLM sessions. Addressing the "Privacy Paradox" where users assume "no login" ensures anonymity, UMA constructs a robust behavioral fingerprint driven by two core technical innovations. This tool establishes a Multi-aspect Feature Space by integrating Content, Stylometrics, Interaction, and Personality. Unlike traditional methods that rely on content matching, this multi-dimensional approach ensures robust tracking even when users drastically switch topics or disguise their writing style.
UMA introduces an Uncertainty-Aware Fusion mechanism. This algorithm mathematically models the reliability of each feature dimension, dynamically suppressing the noise inherent in short, sparse texts to prevent over-fitting. Beyond these core technologies, the system also identifies the "Mirroring Effect," utilizing the AI’s own responses as a secondary signal source. Validated on the WildAuth benchmark (1,755 real-world dialogues), UMA achieves a state-of-the-art 92.05% AUC, significantly outperforming existing baselines. It serves as a critical infrastructure tool for digital forensics and AI safety.
