SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing
Researchers have introduced Statistical Membership Inference (SMI), a novel framework for auditing machine unlearning processes. Traditional methods using Membership Inference Attacks (MIAs) often overestimate unlearning effectiveness due to an alignment bias, where unlearned samples differ from non-member samples in ways that mislead MIA. SMI offers a training-free approach that reformulates auditing as estimating the non-member mixture proportion in the unlearned feature distribution, providing a more reliable and efficient alternative with theoretical guarantees and strong empirical results. AI
IMPACT Introduces a more reliable and efficient method for auditing machine unlearning, potentially improving data privacy in AI systems.