Nicolas Sournac

Robustness Meets Uncertainty: Evidential Adversarial Training for Robust Selective Classification

Nicolas Sournac, Ahmed Baha Ben Jmaa, Bertrand Braeckeveldt

Safety-critical applications require classifiers that are both robust and reliable. Adversarial training is a widely adopted defense for improving robustness in deep neural networks; however, its effect on the reliability of predictive uncertainty remains underexplored. We invest…Read more