Evaluating tensor-related metrics for early fault detection in CNNs

Today's electronic devices are increasingly powered by deep learning algorithms to perform various tasks, from pedestrian recognition in self-driving cars to detecting health problems in low-cost, low-power wearable devices. However, despite the claimed built-in redundancy of deep learning models, the literature shows that they are susceptible to random-hardware faults: even a single corrupted bit can lead to catastrophic failures. Therefore, it is crucial to find out novel, smart, and low-cost fault detection solutions that can be used in the field to stop the propagation of critical faults. This study explores the fault-detection capabilities of mathematical metrics typically used in domains such as image processing, audio analysis, and regression. These metrics, applied to intermediate output tensors of convolutional layers, effectively detect early failures and stop their propagation, saving computational time, costs, and power. Fault injection campaigns are performed on three Convolutional Neural Networks (CNNs) to test the fault detection capability of 15 metrics. The outcomes show their effectiveness in detecting permanent faults with a very high true positive rate. Among the different metrics, the experimental analysis highlighted three metrics that demonstrated notable performance based on their Receiver Operating Characteristic (ROC) curves and corresponding confusion matrices: Minkowski distance, Mean Squared Error (MSE), and the TSum metric. By introducing trade-offs between sensitivity (true positive rate) and specificity (false-positive rate), it is possible to identify between 88.63% and 97.13% of critical faults while sacrificing only between 0.27% and 0.31% of total inferences which are unnecessarily re-executed.