Sequential Auditing for f-Differential Privacy
Researchers have developed new auditors to empirically assess the Differential Privacy (DP) of algorithms, focusing on the expressive $f$-DP concept. These auditors can detect privacy violations across the full privacy spectrum with statistical significance, without requiring a pre-specified sample size. The method adaptively determines the optimal number of samples, significantly reducing the sampling cost, which is particularly beneficial for expensive training procedures like DP-SGD. The auditors support both whitebox and blackbox settings and can be integrated into one-run frameworks. AI
IMPACT This research could lead to more efficient and effective privacy auditing for AI models, particularly those trained with methods like DP-SGD.