Early Detection of Permanent Faults in DNNs Through the Application of Tensor-Related Metrics

Computational models based on deep learning are today integrated in many safety-critical domains. These algorithms, such as deep neural networks (DNNs), are rapidly growing in size, reaching billions or even trillions of parameters. This factor brings big challenges not only for performance goals but also for dependability aspects such as reliability. The larger the model, the more challenging the reliability assessment becomes. It is now crucial to develop new test approaches supported by acceptable computational costs for the detection of random-hardware faults such as permanent faults, which may change the predictions of DNNs. The aim of this paper is to leverage tensor-related metrics to early detect faulty behaviors during the inference of DNNs. This involves calculating metrics applied to tensors across various domains (such as image processing, audio analysis, and regression) on the Output Feature Maps (OFMs) of a layer. This analysis allows knowing in advance the effect that a permanent fault will have on the output of the DNN application. The effectiveness of the approach has been experimentally demonstrated by means of software fault injection campaigns considering faults affecting weights of Convolutional Neural Networks (CNNs), i.e., ResNet20 and MobileNetV2. The quality of the metrics is discussed in terms of the trade-off between energy consumption and the ability to differentiate between critical and non-critical faults.