Reproduce “Low-Cost High-Power Membership Inference Attacks” on OpenResearch

About this paper

Sajjad Zarifzadeh, Philippe Liu, Reza Shokri

Researchers at the National University of Singapore developed the Robust Membership Inference Attack (RMIA), a statistical test that substantially increases attack power and efficiency in identifying training data points in machine learning models. RMIA demonstrates superior performance with limited computational resources and enhanced robustness across various data distributions and model architectures.

Key insights

• Many prior strong membership inference attacks can be viewed as simplifications of RMIA, often lacking proper calibration or specific terms, which explains their observed instabilities and performance gaps.
• The core of achieving high attack power and robustness lies in finely modeling the null hypothesis by contrasting a target data point against a representative set of random samples from the population.
• Practical and scalable privacy auditing is feasible through computationally efficient attacks that maintain high power and robustness with limited reference models, under diverse real-world constraints.

Summary by alphaXiv.