Opportunities and Challenges in Deep Learning Adversarial Robustness: A Survey

As we seek to deploy machine learning models beyond virtual and controlled\ndomains, it is critical to analyze not only the accuracy or the fact that it\nworks most of the time, but if such a model is truly robust and reliable. This\npaper studies strategies to implement adversary robustly trained algorithms\ntowards guaranteeing safety in machine learning algorithms. We provide a\ntaxonomy to classify adversarial attacks and defenses, formulate the Robust\nOptimization problem in a min-max setting and divide it into 3 subcategories,\nnamely: Adversarial (re)Training, Regularization Approach, and Certified\nDefenses. We survey the most recent and important results in adversarial\nexample generation, defense mechanisms with adversarial (re)Training as their\nmain defense against perturbations. We also survey mothods that add\nregularization terms that change the behavior of the gradient, making it harder\nfor attackers to achieve their objective. Alternatively, we've surveyed methods\nwhich formally derive certificates of robustness by exactly solving the\noptimization problem or by approximations using upper or lower bounds. In\naddition, we discuss the challenges faced by most of the recent algorithms\npresenting future research perspectives.\n