CoG-Guided Correction Enhances DNN Fault Tolerance Significantly

Deep Neural Networks (DNNs) deployed in safety-critical environments are susceptible to hardware and memory faults that can corrupt network weights, leading to reliability degradation. This paper introduces a novel Center of Gravity (CoG) guided weight correction method designed to restore these faulty weights. The approach leverages the spatial characteristics of weights within each layer to detect and correct errors using distance-aware rules. A key advantage is that this method requires no retraining of the model or modifications to its architecture, making it highly adaptable. The effectiveness of the CoG-guided correction was rigorously evaluated through fault injection experiments at various Bit Error Rates (BERs). For safety-critical LSTM-based networks, such as StageNet for disease progression tracking and MTFNet for cardiac anomaly detection, the method demonstrated fault tolerance improvements of up to 230x and 6.41x, respectively, at a BER of 10^-3, all while maintaining negligible accuracy loss. When applied to Convolutional Neural Networks (CNNs) like ResNet-18 and VGG-16, improvements of up to 49.55x and 20.79x were observed under similar fault conditions. This work marks the first application of the CoG concept to neural network weight tensors for enhancing model reliability.