Enhancing Deep Neural Network Resilience Through Integrated Reliability-Aware Design and Fault-Aware Training

Deep neural networks (DNNs) are widely used for mission-critical applications but they are vulnerable to transient hardware faults that corrupt inferences without triggering system-level error detection mechanisms. Published radiation experiments show that a single transient fault to a GPU's warp scheduler can corrupt all 32 threads of the warp, changing the feature map going to the next layer and causing a misprediction. Although some counters based on custom silicon or important latency exist, we introduce ResilientNet, a reliability-aware DNN framework to harden DNNs from radiation by (1) using ReLU6 instead of ReLU to clip corruptions, (2) layering Conv->ReLU6->BN instead of Conv->BN->ReLU6 to prevent batch normalization from increasing corrupted statistics, (3) filtering for NaNs to suppress arithmetic poison, and (4) introducing a curriculum that runs the network on experimentally observed fault patterns. The framework is evaluated with 14,600 fault injection experiments with six incrementally more hardened configurations and seven fault models. It is shown that ResilientNet reduces the critical SDC rate from NaNs from 92.0% (95% CI 85.0% to 95.9%) to 17.0% (95% CI 10.9% to 25.6%), a 5.4× reduction. No critical SDC was found for 100 iterations per configuration (95% Wilson upper bound 3.70%). We assessed the classification accuracy and the model's fault tolerance by comparing its performance with that of the counterpart unprotected baseline instance (based on 96.67% and 92.50% accuracy, respectively). The ROC curve, confusion matrices, layer-wise sensitivity heatmaps were assessed. All rates are given with Wilson score confidence intervals. Bounded activation and layer reordering require no additional arithmetic. NaN filtering requires 576 exponent-field comparisons at inference time or 0.35% of this network's arithmetic operations. All fault injection is performed at the application layer on a proof-ofconcept data set. The statements in this section about faults induced by physical radiation environments are based on the cited literature rather than on any experiment conducted in this work.