Go to ICLR 2025 Conference homepageWatch Out!! Your Confidence Might be a Reason for Vulnerability
The tremendous success of deep neural networks (DNNs) in solving `any' complex computer vision task leaves no stone unturned for their deployment in the physical world. However, the concerns arise when natural adversarial corruptions might perturb the physical world in unconstrained images. It is widely known that these corruptions are inherently present in the environment and can fool DNNs. While the literature aims to provide safety to DNNs against these natural corruptions they have developed two forms of defenses: (i) detection of corrupted images and (ii) mitigation of corruptions. So far, very little work has been done to understand the reason behind the vulnerabilities of DNNs against such corruption. We assert that network confidence is an essential component and ask whether the higher it is, the better the decision of a network is or not. Moreover, we ask the question of whether this confidence itself is a reason for their vulnerability against corruption. We extensively study the correlation between the confidence of a model and its robustness in handling corruption. Through extensive experimental evaluation using multiple datasets and models, we found a significant connection between the confidence and robustness of a network.