OpRA: Optimizing Resiliency Assessment for Deep Neural Networks

Deep neural networks (DNNs) and especially convolutional neural networks (CNNs) are increasingly deployed in safety-critical domains, where rigorous reliability assessment is essential. Fault injection (FI) remains a widely adopted technique to evaluate DNN robustness under hardware faults. However, the high computational cost of FI campaigns limits their scalability and practical adoption. These campaigns traditionally rely on three components: the DNN model, a fault list, and an input stimulus set. While fault selection strategies have been extensively studied, the impact of input stimulus selection remains largely underexplored. This work addresses this gap by investigating the role of input stimuli in activating critical faults during FI-based reliability evaluation. We propose a strategy to prioritize inputs that are more likely to expose fault-induced vulnerabilities, guided by uncertainty-based metrics. Our methodology is validated across two fault models: memory-level bit flips and stuck-at faults in a systolic array (SA) accelerator. Results show that uncertainty-ranked inputs significantly increase fault activation and detection efficiency, enabling more focused and cost-effective reliability analysis.