Exploring misclassifications of robust neural networks to enhance adversarial attacks

Progress in making neural networks more robust against adversarial attacks is mostly marginal, despite the great efforts of the research community. Moreover, the robustness evaluation is often imprecise, making it challenging to identify promising approaches. We do an observational study on the classification decisions of 19 different state-of-the-art neural networks trained to be robust against adversarial attacks. This analysis gives a new indication of the limits of the robustness of current models on a common benchmark. In addition, our findings suggest that current untargeted adversarial attacks induce misclassification toward only a limited amount of different classes. Similarly, we find that previous attacks under-explore the perturbation space during optimization. This leads to unsuccessful attacks for samples where the initial gradient direction is not a good approximation of the final adversarial perturbation direction. Additionally, we observe that both over- and under-confidence in model predictions result in an inaccurate assessment of model robustness. Based on these observations, we propose a novel loss function for adversarial attacks that consistently improves their efficiency and success rate compared to prior attacks for all 30 analyzed models.