Reproduce “Curvature Clues: Decoding Deep Learning Privacy with Input Loss Curvature” 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
Deepak Ravikumar, Efstathia Soufleri, Kaushik Roy
This paper presents a novel black-box membership inference attack (MIA) that utilizes input loss curvature as a distinguishing signal for deep learning models. The approach demonstrates superior performance over existing state-of-the-art MIAs on large-scale datasets and provides a theoretical framework linking curvature, privacy, and training set size.
Key insights
- Input loss curvature is consistently higher for test examples than for training examples, providing a distinct signal for membership inference.
- The theoretical upper bound on train-test distinguishability via input loss curvature is shown to be dependent on the training set size (m) in addition to the differential privacy parameter (ϵ).
- Models trained on low-curvature coresets, despite potentially memorizing less, exhibit increased susceptibility to curvature-based membership inference attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.