Revisiting adversarial training for the worst-performing class
This work tackles class imbalance in adversarial training, which is particularly important for safety-critical applications where worst-case performance matters.
The paper addresses the substantial accuracy gap between best and worst-performing classes in adversarial training, proposing a method that explicitly optimizes for the worst class to reduce this disparity. On CIFAR10, it improves worst class accuracy from 23% to 32% while maintaining minimal computational overhead.
Despite progress in adversarial training (AT), there is a substantial gap between the top-performing and worst-performing classes in many datasets. For example, on CIFAR10, the accuracies for the best and worst classes are 74% and 23%, respectively. We argue that this gap can be reduced by explicitly optimizing for the worst-performing class, resulting in a min-max-max optimization formulation. Our method, called class focused online learning (CFOL), includes high probability convergence guarantees for the worst class loss and can be easily integrated into existing training setups with minimal computational overhead. We demonstrate an improvement to 32% in the worst class accuracy on CIFAR10, and we observe consistent behavior across CIFAR100 and STL10. Our study highlights the importance of moving beyond average accuracy, which is particularly important in safety-critical applications.