Adversarial and Out-of-Distribution Perspectives on Deep Neural Network Robustness

Abstract

Deep neural networks (DNNs) have achieved significant success across a broad spectrum of applications, such as autonomous driving, medical imaging, and natural language processing. However, their susceptibility to distributional shifts, including adversarial examples and out-of-distribution (OOD) data, poses serious challenges to their robustness and reliability in safety-critical applications. This thesis examines adversarial and OOD perspectives on DNN robustness, proposing novel approaches to identify adversarial vulnerabilities and strengthen OOD robustness, ultimately improving their reliability in real-world applications.

We begin by surveying the intersection of adversarial robustness and OOD detection, establishing a taxonomy centered around distributional shifts. Next, we investigate the adversarial vulnerability of inputs, proposing a novel metric based on the clipped gradients of the loss with respect to the input. This metric reveals that some inputs are inherently more susceptible to adversarial perturbations, a property that can be leveraged to improve black-box attack pipelines.

To improve OOD detection, we propose two algorithms that utilise class rank information implicitly learned by DNNs during standard training with cross-entropy. These methods are motivated by the observation that class ranking patterns in in-distribution (ID) data are more consistent and deterministic than the stochastic patterns seen in OOD data. Accordingly, first, we introduce ExCeL, a post-hoc detector that integrates extreme and collective information from the output layer of DNNs. By combining the maximum logit (i.e., extreme information) with novel rank-based score (i.e., collective information), ExCeL achieves consistent and competitive performance across diverse OOD scenarios. Finally, we present CRAFT, a fine-tuning approach that further strengthens OOD robustness by optimising DNNs based on the implicit class ranking information learned during pre-training.