Dealing with Robustness of Convolutional Neural Networks for Image Classification

SW-based systems depend more and more on AI also for critical tasks. For instance, the use of machine learning, especially for image recognition, is increasing ever more. As state-of-the-art, Convolutional Neural Networks (CNNs) are the most adopted techniques for image classification. Although they are proved to have optimal results, it is not clear what happens when unforeseen modifications during the image acquisition and elaboration occur. Thus, it is very important to assess the robustness of a CNN, especially when it is used in a safety critical system, as, e.g., in the medical domain or in automated driving systems. Most of the analyses made about the robustness of CNNs are focused on adversarial examples which are created by exploiting the CNN internal structure; however, these are not the only problems we can encounter with CNNs and, moreover, they may be unlikely in some fields. This is why, in this paper, we focus on the robustness analysis when plausible alterations caused by an error during the acquisition of the input images occur. We give a novel definition of robustness w.r.t. possible input alterations for a CNN and we propose a framework to compute it. Moreover, we analyse four methods (data augmentation, limited data augmentation, network parallelization, and limited network parallelization) which can be used to improve the robustness of a CNN for image classification. Analyses are conducted over a dataset of histologic images.