Towards Understanding the Dynamics of the First-Order Adversaries
ICML, pp. 2484-2493, 2020.
We present our main results on the convergence of projected gradient descent
An acknowledged weakness of neural networks is their vulnerability to adversarial perturbations to the inputs. To improve the robustness of these models, one of the most popular defense mechanisms is to alternatively maximize the loss over the constrained perturbations (or called adversaries) on the inputs using projected gradient ascen...More
PPT (Upload PPT)