Reproduce “SoK: Challenges in Tabular Membership Inference Attacks” 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
Cristina Pêra, Tânia Carvalho, Maxime Cordy, Luís Antunes
This Systematization of Knowledge paper empirically investigates Membership Inference Attacks (MIAs) on tabular data, offering refined taxonomies for centralized and federated learning. It demonstrates that MIAs are generally less effective on tabular datasets than reported for other data types but uncovers a critical vulnerability for unique records (single-outs) and reveals that attacks can transfer across different model architectures in black-box settings.
Key insights
- MIAs are generally less effective on tabular data compared to other data types, and high overfitting is a sufficient, but not strictly necessary, condition for attack success.
- "Single-outs" (unique records) exhibit high vulnerability to MIAs, irrespective of the overall low attack efficacy on the full dataset, emphasizing the need for robust data anonymization.
- MIAs demonstrate transferability across mismatched model architectures, enabling effective black-box attacks even without precise knowledge of the target model's internals.
- Model hardening techniques and specific Federated Learning configurations can effectively mitigate MIA efficacy, though sometimes with an observed privacy-utility trade-off.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.