Fault Tolerance of Adversarially Robust Pruned Models Explored

Deploying deep neural networks (DNNs) on resource-constrained hardware, particularly neuromorphic systems, presents a triple challenge: the need for model compression through pruning, vulnerability to adversarial attacks, and susceptibility to hardware-induced errors like stuck-at-zero faults. This research empirically examined how these three factors interact to influence the reliability of convolutional neural networks (CNNs). Using a compact three-layer CNN on MNIST, the study compared the fault tolerance of naturally and adversarially trained models under various conditions. Key findings indicate that while adversarial training enhances a model's resilience to input perturbations, it paradoxically increases its sensitivity to hardware faults. Counterintuitively, pruning the model did not lead to a significant increase in fault sensitivity, and varying the pruning level had minimal impact across different fault rates and attack strengths. These results underscore the importance of considering both adversarial robustness and hardware reliability jointly during model design and deployment, rather than in isolation.