Adversarial learning aims to identify weaknesses in machine learning models. The goal is to identify potential problems that cannot be found using traditional evaluation using test sets. It has been used successfully in a wide range of applications, typically focused on a specific model or domain. In image classification, methods have been developed to fool models that recognize traffic signs by rather simple modifications of pictures. Another direction of adversarial learning aims to identify examples that could break or improve the training of the model if that example would be added to the training.

Our research is focusing on defending machine learning models from adversarial attacks. No defense is completely immune to adversarial examples. Most existing approaches are tied to a specific machine learning model and struggle to preserve good performance on clean inputs. As a result, few adversarial defenses have been deployed in real-world applications. Additionally, we are interested in the implications of adversarial machine learning on developing machine learning models, their training and evaluation, and their application.