As AI models process larger datasets with sensitive information, verifying machine unlearning has become critical for privacy compliance. Current statistical methods like two-sample testing struggle with high-dimensional data and produce false positives when comparing models trained differently. Google AI Blog researchers introduce Regularized f-Divergence Kernel Tests at AISTATS 2026 to address these limitations.

The framework uses f-divergences including Hockey-stick divergence to measure whether unlearned models are closer to safely retrained versions or original compromised ones. Unlike traditional approaches that require manual tuning of kernel parameters, this method automatically selects optimal configurations while controlling false positives for any sample size. The approach leverages kernel regularization to handle complex optimization problems efficiently.

Testing on synthetic benchmarks and high-energy physics data demonstrated superior performance over baseline methods. For privacy auditing, the framework successfully identified violations in non-private mechanisms while catching subtle anomalies that standard tests missed. Different f-divergences act as specialized sensors for various types of data shifts.

Researchers validated their approach across multiple unlearning algorithms including Selective Synaptic Dampening and pruning techniques. The relative distance testing method provides a practical solution for organizations needing to verify compliance without computationally expensive full retraining.