Risk Analysis of One-Pixel Image Defects in Safety-Critical Deep Neural Networks

Deep neural networks (DNNs) are widely considered essential for developing perception systems in autonomous applications. These models are often vulnerable to small perturbations in input data, even if the changes appear negligible to a human observer. This vulnerability introduces an additional risk of failure in safety-critical systems during normal operation. Unfortunately, there is currently no quantitative risk analysis addressing such image defects. In contrast, this work examines the risk that one-pixel defects may occur naturally within image data, and evaluates how frequently such seemingly minor defects can lead to incorrect decisions by neural networks. Extensive experiments reveal that the number of impactful image defects may be relatively high, depending on both the DNN architecture and the dataset used. These findings establish that image defects require significant attention and it might not be sufficient to argue for an acceptable level of safety based solely on the low probability of occurrence these defects.