Reproduce “Reconstruction and Membership Inference Attacks against Generative Models” on OpenResearch

About this paper

Benjamin Hilprecht, Martin Härterich, Daniel Bernau

Researchers at TU Darmstadt and SAP SE introduced Monte Carlo and Reconstruction membership inference attacks, demonstrating superior performance against generative models and revealing Variational Autoencoders are significantly more susceptible to information leakage than Generative Adversarial Networks. The work also formalized and validated a highly effective set membership inference attack for auditing the unauthorized use of specific datasets.

Key insights

• Generative models, particularly VAEs, are highly susceptible to membership inference attacks, with VAEs demonstrating greater vulnerability compared to GANs under similar training conditions.
• The success rates of membership inference attacks can serve as a quantitative indicator of overfitting and generalization capabilities in generative models, offering a proxy for model quality in MLaaS scenarios.
• Set Membership Inference is highly effective for regulatory actors, achieving high accuracies that validate its practical utility for auditing and ensuring compliance with data privacy standards.

Summary by alphaXiv.