Reproduce “Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable” 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
Martin Bertran, Shuai Tang, Michael Kearns, Jamie Morgenstern, Aaron Roth, Zhiwei Steven Wu
The paper demonstrates that machine unlearning, even through full retraining, introduces a vulnerability enabling high-accuracy reconstruction of deleted data, particularly for simple models like linear regression. It shows that observing model parameters before and after deletion allows an attacker to recover private training samples, challenging the assumption that simple models pose low privacy risk.
Key insights
- The act of machine unlearning itself can introduce significant privacy risks, even for simple models previously thought to be secure, by exposing information about deleted data.
- Achieving functional equivalence (a model state as if data was never there) in unlearning is insufficient for privacy; the process of updating the model can be highly informative to an attacker.
- The findings underscore the necessity of robust privacy-preserving mechanisms, like differential privacy, for the unlearning process itself, extending their scope beyond initial model training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.