Similarity-driven adversarial testing of neural networks

Katarzyna Filus, Joanna Domanska

Abstract: Highlights•Novel methods for similarity-driven label choosing for adversarial attacks.•Deep networks’ testing based on the worst/best-case scenarios.•Similarity of target labels is assessed from the network’s/semantic perspective.•Testing in a more effective, transparent and human-centric manner.•Contribution to explainable testing, standardization, defining security requirements.•Superior results compared to common strategies.