Universität Bielefeld Play

[BA/MA/Project]

XAI Methods on Adversarial Inputs

Contact: André Artelt

Many XAI methods for deep learning based vision systems (e.g. image classifier) are based on saliency maps — i.e. highlighting parts of the image that were relevant for computed predictions. At the same time, adversarial attacks still pose a threat to such deep learning models.

The aim of potential BA/MA theses or projects is to investigate saliency map based XAI methods on adversarials — i.e. how (if at all) does an adversarial attack manifests itself in the generated explanations. The exact topic/objective will be tailored toward the student’s interest and the required number of credit points.

Keywords: Explainable AI, Deep Learning, Adversarial Attacks

Literature

  1. https://arxiv.org/pdf/2107.01943.pdf