A New Online Fault Detection Mechanism for Neural Network Applications

This paper proposes a novel strategy for detecting hardware-induced misbehaviors during the execution of Convolutional Neural Networks (CNNs). The fault detection strategy relies on an auxiliary lightweight Multi-Layer Perceptron (MLP) acting on the outputs of the first layer of the monitored CNN as an early exit instrument, and enabling the detection of faults possibly affecting the CNN computations. Specifically, the outcomes of the CNN and the MLP are compared across a sequence of observations (observation window), calculating the level of agreement between them. As a result, an incorrect prediction from a CNN caused by a hardware fault can be identified when the agreement drops below a set threshold. This mechanism was implemented and evaluated using a set of representative CNNs and datasets. The results showed that the system can detect between 78.60% to 97.79% of faults, depending on the CNN and the dataset.