Adversarial Examples in Processing of High-Tech Imaging Using DNN

Abstract

The article investigates the author-detected effect of noise from devices forming reconstructed high-tech images on their classification performance by means of deep neural networks. The main observation is that Computed Tomography and Magnetic Resonance Imaging devices introduce noise, which causes an effect called “natural adversarial examples”, which is similar to the adversarial examples attack. These noisy images alter a manifold in the intermediate feature space of a deep neural network, making the network less robust. The article suggests simulating their effects using the fast gradient sign method (FGSM) attack model. To mitigate their influence, it is proposed to augment the training dataset with noises of a similar nature as natural adversarial examples (AEs) (rather than trying to remove the latter from the original dataset). The paper describes a training pipeline and its parametrization for achieving best performance and robustness to the natural adversarial examples for high-tech medical images classification using deep neural networks (DNN). As a result of the research, we have shown that due to the successful selection of the architecture and network parameters, primarily the variants of the activation function and augmentation of the original dataset, it is possible to achieve a suboptimal ratio of accuracy and f1-score, on the one hand, and robustness to natural AE, on the other hand.