Robust convolutional neural network with integrated multiscale attention mechanism against adversarial attacks

Abstract

Convolutional neural networks have been widely used in recent years for image analysis in areas such as tumor segmentation and disease detection. However, a growing body of research has shown that applying small disturbances to an image can wreak havoc on the results quite differently from a correct diagnosis. This is of great concern to organizations that have already deployed AI medical detection systems, and related research is proceeding at a rapid pace. In this paper, we investigate different ways of attacking medical images against adversarial attacks, and the reasons why convolutional neural networks show performance degradation against perturbation noise are analyzed. We propose a new method to improve the robustness of the convolutional neural network classification system by adding attention modules at different locations of the model to improve the performance of the model's feature extraction as a way to improve robustness. Finally, the performance of the network model is examined using the most popular attacks available. We validate on a medical image dataset and demonstrate that the proposed method allows the convolutional neural network to successfully defend against attacks on counter-attack samples without significant performance degradation. Our study suggests that these findings may provide some positive effects on future medical image analysis and processing as well as exploring deep image features.