Implications of Solution Patterns on Adversarial Robustness

Abstract

Empirical robustness evaluation (RE) of deep learning models against adversarial perturbations involves solving non-trivial constrained optimization problems. Recent work has shown that these RE problems can be reliably solved by a general-purpose constrained-optimization solver, PyGRANSO with Constraint-Folding (PWCF). In this paper, we take advantage of PWCF and other existing numerical RE algorithms to explore distinct solution patterns in solving RE problems with various combinations of losses, perturbation models, and optimization algorithms. We then provide extensive discussions on the implications of these patterns on current robustness evaluation and adversarial training. A comprehensive version of this work can be found in [19].