Bayesian Sampling Boosts Membership Inference Attacks
A new technique called Bayesian Membership Inference Attack (BMIA) has been introduced by researchers, leveraging Bayesian sampling to enhance both the efficacy and efficiency of membership inference attacks on machine learning systems. Conventional membership inference attacks necessitate the training of several reference models to gauge conditional score distributions, leading to significant computational expenses. In contrast, BMIA utilizes Laplace approximation on a single reference model to derive a posterior distribution over model parameters, allowing for direct estimation of the conditional score distribution. Theoretically, Bayesian sampling minimizes intra-model variance, boosting the attack's effectiveness. This concept inspires a multi-reference variant that further optimizes performance with the availability of extra reference models. This research is available on arXiv under ID 2503.07482.
Key facts
- Membership Inference Attacks (MIAs) estimate whether a data point was used in training a model.
- Existing state-of-the-art MIAs rely on training multiple reference models, causing high computational overhead.
- BMIA uses Bayesian sampling with Laplace approximation on a single reference model.
- Bayesian sampling reduces intra-model variance, improving attack power.
- A multi-reference variant of BMIA further enhances performance when additional models are available.
- The paper is available on arXiv with ID 2503.07482.
- The method is called Bayesian Membership Inference Attack (BMIA).
- The approach enables direct estimation of conditional score distribution.
Entities
Institutions
- arXiv