How does Bayesian Sampling help Membership Inference Attacks?

Name: How does Bayesian Sampling help Membership Inference Attacks?
Uploaded: 2026-10-22T15:26:42.000268+00:00
Duration: 899 s
Description: Bayesian Membership Inference Attack (BMIA) is a cost-efficient method for conditional membership inference using Bayesian sampling and Laplace approximation on a single reference model. It estimates the conditional score distribution by sampling weights from an approximated posterior, proving that this approach reduces intra-model variance to improve attack power.

Bayesian Membership Inference Attack (BMIA) is a cost-efficient method for conditional membership inference using Bayesian sampling and Laplace approximation on a single reference model. It estimates the conditional score distribution by sampling weights from an approximated posterior, proving that this approach reduces intra-model variance to improve attack power.

Open the full episode