Jacobian Regularization for Mitigating Universal Adversarial Perturbations

Universal Adversarial Perturbations (UAPs) are input perturbations that can fool a neural network on large sets of data. They class attacks represents significant threat as they facilitate realistic, practical, and low-cost networks. In this work, we derive upper bounds for the effectiveness UAPs based norms data-dependent Jacobians. We empirically verify Jacobian regularization greatly increases model robustness to by up four times whilst maintaining clean performance. Our theoretical analysis also allows us formulate metric strength shared adversarial between pairs inputs. apply benchmark datasets show it is highly correlated with actual observed robustness. This suggests realistic practical universal be reliably mitigated without sacrificing accuracy, which shows promise machine learning systems.