Sequentially Auditing Differential Privacy

A new practical sequential test for auditing differential privacy guarantees of black-box mechanisms has been proposed. This test processes streams of outputs, allowing for anytime-valid inference while controlling Type I error. It significantly reduces the sample size needed for detecting violations from 50,000 to just a few hundred examples across various mechanisms. Notably, it can identify DP-SGD privacy violations in under one training run, unlike previous methods that required complete model training.