Peter Lorenz

Peter Lorenz, Mario Fernandez, Jens Müller et al.

Detecting out-of-distribution (OOD) inputs is critical for safely deploying deep learning models in real-world scenarios. In recent years, many OOD detectors have been developed, and even the benchmarking has been standardized, i.e. OpenOOD. The number of post-hoc detectors is growing fast. They are showing an option to protect a pre-trained classifier against natural distribution shifts and claim to be ready for real-world scenarios. However, its effectiveness in dealing with adversarial examples (AdEx) has been neglected in most studies. In cases where an OOD detector includes AdEx in its experiments, the lack of uniform parameters for AdEx makes it difficult to accurately evaluate the performance of the OOD detector. This paper investigates the adversarial robustness of 16 post-hoc detectors against various evasion attacks. It also discusses a roadmap for adversarial defense in OOD detectors that would help adversarial robustness. We believe that level 1 (AdEx on a unified dataset) should be added to any OOD detector to see the limitations. The last level in the roadmap (defense against adaptive attacks) we added for integrity from an adversarial machine learning (AML) point of view, which we do not believe is the ultimate goal for OOD detectors.

Peter Lorenz, Dominik Strassel, Margret Keuper et al.

Recently, RobustBench (Croce et al. 2020) has become a widely recognized benchmark for the adversarial robustness of image classification networks. In its most commonly reported sub-task, RobustBench evaluates and ranks the adversarial robustness of trained neural networks on CIFAR10 under AutoAttack (Croce and Hein 2020b) with l-inf perturbations limited to eps = 8/255. With leading scores of the currently best performing models of around 60% of the baseline, it is fair to characterize this benchmark to be quite challenging. Despite its general acceptance in recent literature, we aim to foster discussion about the suitability of RobustBench as a key indicator for robustness which could be generalized to practical applications. Our line of argumentation against this is two-fold and supported by excessive experiments presented in this paper: We argue that I) the alternation of data by AutoAttack with l-inf, eps = 8/255 is unrealistically strong, resulting in close to perfect detection rates of adversarial samples even by simple detection algorithms and human observers. We also show that other attack methods are much harder to detect while achieving similar success rates. II) That results on low-resolution data sets like CIFAR10 do not generalize well to higher resolution images as gradient-based attacks appear to become even more detectable with increasing resolutions.