Adversaries meet PDEs: A First Look at Fluidic Adversarial Robustification

January - April, 2019, Georgia Institute of Technology (ECE 6560: PDEs for Image Processing & Computer Vision)

Invented a Lagrangian to perform blind robustification of adversarial examples, and proposed a stable discretization scheme for its gradient flow. The proposed blind approach targets to destroy the structure of adversarial perturbation while causing minimal damage to the information signal.