Improving the Tightness of Convex Relaxation Bounds for Training Certifiably Robust Classifiers
This work addresses the problem of enhancing certifiable robustness against adversarial attacks for machine learning practitioners, though it is incremental as it builds on existing convex relaxation methods.
The paper tackled the gap between certifiable and empirical robustness in neural networks by proposing two regularizers that improve the tightness of convex relaxation bounds, resulting in higher certified accuracy across all experiments.
Convex relaxations are effective for training and certifying neural networks against norm-bounded adversarial attacks, but they leave a large gap between certifiable and empirical robustness. In principle, convex relaxation can provide tight bounds if the solution to the relaxed problem is feasible for the original non-convex problem. We propose two regularizers that can be used to train neural networks that yield tighter convex relaxation bounds for robustness. In all of our experiments, the proposed regularizers result in higher certified accuracy than non-regularized baselines.