Revisiting adversarial training for the worst-performing class

TL;DR

This paper introduces CFOL, a new adversarial training method that explicitly optimizes for the worst-performing class, significantly reducing class accuracy gaps and improving robustness in datasets like CIFAR10.

Contribution

The paper proposes a novel min-max-max optimization framework called CFOL that explicitly targets the worst class in adversarial training, with proven convergence guarantees and minimal computational overhead.

Findings

Improved worst class accuracy to 32% on CIFAR10

Consistent improvements across CIFAR100 and STL10

Highlights importance of focusing on worst-case class performance

Abstract

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.