Robustness Should Not Be at Odds with Accuracy - A Statistical Learning Theory Perspective

Watch a 19-minute conference talk from the 2022 Symposium on Foundations of Responsible Computing where York University's Ruth Urner explores the relationship between adversarial robustness and accuracy in deep learning models. Examine how imperceptible perturbations can cause neural networks to make incorrect predictions, and discover why traditional approaches to adversarial robustness may conflict with accuracy requirements. Learn about a novel approach using locally adaptive robust loss that maintains consistency in 1-nearest neighbor classification under deterministic labels. Follow the systematic analysis of standard and robust Bayes classifiers, understand data-informed adaptive robustness radius, and explore how proper modeling of adversarial examples should align with underlying data-generating processes. Gain insights into why robustness and accuracy should be complementary rather than conflicting goals in machine learning systems.