For models to achieve adversarial robustness, they need to perform under small perturbations crafted to mislead them, which is a key part of the robustness axis.
The tremendous progress and success of Deep Neural Networks has also come with discovery of many of their weaknesses. Specifically, they have been proved to be vulnerable to small/insignificant perturbations that would be imperceptible to humans but that easily misleads models. These so-called adversarial examples are a threat to the deployment of these models in safety critical applications. The goal of this project is to be able to detect and/or to achieve adversarial robustness to face these kinds of instances.
Team