Uncovering hidden vulnerabilities in CNNs through evolutionary-based Image Test Libraries

Over the past years, Convolutional Neural Networks (CNNs) have become highly prominent in diverse artificial intelligence applications, proving remarkable performance across numerous tasks. Nonetheless, as CNNs find greater integration in safety-critical domains, the need to ensure their safety and reliability has emerged as a crucial issue. Efforts are underway to establish safety standards that specifically address the unique challenges posed by CNNs. As an example, these standards emphasize the importance of detecting in field permanent faults that may occur during the operational phase of devices. However, it is worth noting that CNNs have an intrinsic masking ability that complicates the in-field detection of permanent faults. This peculiar feature is a result of the distributed and parallel structure of artificial neural networks, that can stop the propagation of some hardware-induced faults by masking their effects. The aim of this research is to develop a software-based testing solution that can effectively reduce the masking effect of CNNs. This will be achieved by generating Image Test Libraries (ITLs) specifically designed to expose, when processed by a fully configured CNN, hard-to-detect faults and ensure their propagation to the output of the CNN. The proposed solution intends to uncover hidden vulnerabilities and enhance the overall reliability of CNNs through targeted testing strategies. An evolutionary algorithm is exploited to generate ITLs able to effectively detect the occurrence of faults at the output of the neural network. Results gathered from experiments targeting permanent faults on weights using ResNet-18, ResNet-34, and DenseNet-161 demonstrate the suitability of the proposed approach.