Bluff: Interactively Deciphering Adversarial Attacks on Deep Neural Networks

Deep neural networks (DNNs) are now commonly used in many domains. However,\nthey are vulnerable to adversarial attacks: carefully crafted perturbations on\ndata inputs that can fool a model into making incorrect predictions. Despite\nsignificant research on developing DNN attack and defense techniques, people\nstill lack an understanding of how such attacks penetrate a model's internals.\nWe present Bluff, an interactive system for visualizing, characterizing, and\ndeciphering adversarial attacks on vision-based neural networks. Bluff allows\npeople to flexibly visualize and compare the activation pathways for benign and\nattacked images, revealing mechanisms that adversarial attacks employ to\ninflict harm on a model. Bluff is open-sourced and runs in modern web browsers.\n