Model Zoo for Robust Models are less Over-Confident

Model Zoo (PyTorch) of non-adversarially trained models for Robust Models are less Over-Confident (NeurIPS'22)

Abstract: "Regardless of the success of convolutional neural networks (CNNs) in many academic benchmarks of computer vision tasks, their application in real-world is still facing fundamental challenges, like the inherent lack of robustness as unveiled by adversarial attacks. These attacks target to manipulate the network's prediction by adding a small amount of noise onto the input. In turn, adversarial training (AT) aims to achieve robustness against such attacks by including adversarial samples in the trainingset. However, a general analysis of the reliability and model calibration of these robust models beyond adversarial robustness is still pending. In this paper, we analyze a variety of adversarially trained models that achieve high robust accuracies when facing state-of-the-art attacks and we show that AT has an interesting side-effect: it leads to models that are significantly less overconfident with their decisions even on clean data than non-robust models. Further, our analysis of robust models shows that not only AT but also the model's building blocks (activation functions and pooling) have a strong influence on the models' confidence."