Universal Adversarial Perturbations

With Alhussein Fawzi; UCLA, DeepMind

Universal Adversarial Perturbations: Fooling Deep Networks with a Single Image

The robustness of classifiers to small perturbations of the data points is a highly desirable property when the classifier is deployed in real and possibly hostile environments. Despite achieving excellent performance on recent visual benchmarks, I will show in this talk that state-of-the-art deep neural networks are highly vulnerable to universal, image-agnostic, perturbations. After demonstrating how such universal perturbations can be constructed, I will analyse the implications of this vulnerability and provide a geometric explanation for the existence of such perturbations via an analysis of the curvature of the decision boundaries.

Add to your calendar or Include in your list