Reproduce “Enhanced Membership Inference Attacks against Machine Learning Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiayuan Ye, Aadyaa Maddi, Sasi Kumar Murakonda, Vincent Bindschaedler, Reza Shokri
A hypothesis testing framework for membership inference attacks (MIAs) against machine learning models is introduced, formalizing existing methods and proposing new black-box attacks to systematically reduce attacker uncertainty. This approach provides more precise privacy risk quantification, with the novel Attack D achieving an AUC of 0.831 on Purchase100 and substantially higher true positive rates at low false positive rates compared to prior techniques.
Key insights
- Systematic reduction of attacker uncertainty, achieved through formal inference games, yields more powerful and interpretable membership inference attacks.
- Attacks leveraging sample-dependent and model-and-sample-dependent information (Attack R and Attack D) produce sharper attack signals and significantly outperform general or shadow model-based approaches.
- Differential analysis indicates that data point vulnerability is influenced by factors beyond general overfitting, such as intrinsically high-loss records and the presence of 'latent neighbors' in the training dataset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.