New framework for auditing machine unlearning

Google Research · 1 month ago

Researchers introduced Regularized f-Divergence Kernel Tests, a new auditing framework presented at AISTATS 2026 for verifying machine unlearning in AI models. The framework uses multiple divergence measures to detect whether unlearned models successfully removed specific training data, with the hockey-stick divergence test detecting privacy violations using only thousands of samples compared to millions required by previous methods. This enables auditors to mathematically prove privacy compliance with minimal manual tuning and fewer data samples, addressing regulatory requirements like GDPR's Right to be Forgotten.