Evidential Adversarial Training Boosts Robustness and Uncertainty.

For safety-critical applications, classifiers must be both robust against adversarial attacks and reliable in their uncertainty predictions. While adversarial training is a common defense for robustness, its impact on the reliability of predictive uncertainty has been largely unexplored. This study addresses this gap by systematically analyzing adversarial robustness alongside selective classification. The researchers established a unified benchmark to evaluate the trade-off between robustness and uncertainty across various settings, including clean, adversarial, and common-corruption scenarios. Their analysis of state-of-the-art adversarial training methods revealed a consistent issue: many approaches improve robust accuracy but degrade uncertainty ranking, leading to suboptimal selective behavior. To counter this, they propose Evidential Adversarial Training (EV-AT). This method models uncertainty using a Dirichlet distribution and integrates two key loss components: an evidence-based loss that promotes clean accuracy and reliable uncertainty, and a robust evidence-alignment loss that matches clean and adversarial predictions in the log Dirichlet-parameter space. Extensive experiments demonstrate that EV-AT significantly shifts the Pareto frontier for robustness-uncertainty trade-offs, surpassing previous adversarial training methods.