Reproduce “Promoting Robustness of Randomized Smoothing: Two Cost-Effective Approaches” 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
Linbo Liu, Trong Nghia Hoang, Lam M. Nguyen, Tsui-Wei Weng
This paper presents AdvMacer and EsbRS, two cost-effective approaches that enhance randomized smoothing for certified adversarial robustness. AdvMacer improves robust training by maximizing certified radius on adversarial examples, while EsbRS leverages theoretically grounded optimal weighting for mixed-model ensembles, leading to superior certified robustness and training efficiency.
Key insights
- Maximizing certified radius on adversarial examples (rather than clean data) during training leads to more robust decision boundaries and improved certified robustness.
- Mixed-model ensembles, combining different types of robustly trained classifiers, can achieve superior certified robustness compared to ensembles of similar models, particularly when logits from different types have lower covariance.
- Theoretically derived optimal weights for ensemble members, based on estimated variances and covariances of model logits, further enhance the certified radius of smoothed classifiers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.