News

After demonstrating their capabilities in everyday tasks, this machine learning models will get increasingly prevalent in critical applications as well. However, their vulnerability to adversarial examples is one of the longest known traits, that raises concerns about their reliability. Building on a previous talk in this seminar, this talk will focus on the creation and analysis of robust adversarial examples for deep neural networks. These kinds of adversarial examples are of particular interest, since they may abstract away from one specific AI model towards a more general class of DNNs. Thus, robust adversarial examples can give insights to the broader limitations of AI approaches and thus help assess the trustworthiness of AI systems. Particularly, in this talk, various methods, from literature and newly designed ones, for generating these adversarial examples will be presented. In this context, we also see, how several notions of adversarial examples exist in different context. Furthermore, will discuss measurements for the robustness of an adversarial example and evaluate the various methods on small examples based on these measurements.