UCLA

In this article we briefly review current research in adversarial attacks and defenses and form a basis for a theoretical explanation as to why a generative energy model is the solution to the defense problem as it exists for securing naturally trained classifiers. We further expand on this topic and discuss future efforts toward the use of a generalized adversarial defense framework based on Stochastic Security to defend against the strongest known adversarial attacks. We further expand on this idea and demonstrate that Energy-based models can be extended towards multiple tasks and datasets. Furthermore, we discuss some architectural improvements to the framework that lead to improvements in synthesis and defense (The Hat-EBM and the Fixer). This work lies at the intersection of generative modeling, adversarial defense, and chaotic dynamics.