The CPG attended the USENIX Security Symposium in Boston on 13-15 August 2025. Igor Shilov presented the paper ‘Free Record-Level Privacy Risk Evaluation Through Artifact-Based Method’, which introduces a novel approach to identifying data points most vulnerable to membership inference attacks (MIAs) without requiring additional computational resources. The method leverages training artifacts such as loss traces, the progression of loss values for individual data points across training epochs, to accurately predict which records are at highest risk of privacy leakage.
