Open Internet by MindsNet

Evaluating Failure Coherence in Adversarial Attacks on Vision Models

Current adversarial robustness benchmarks focus on how easily models can be broken, but do not consider the nature of the failure. When a model misclassifies, does it fail in a semantically related way or in a completely random manner? This challenge aims to differentiate between boundary resistance and failure coherence in vision models under adversarial attacks.

Computing & Technology, Computer Science, Artificial Intelligence