MoAR-CNN: Multi-Objective Adversarially Robust Convolutional Neural Network for SAR Image Classification

Abstract

Deep neural networks (DNNs) have been widely applied to the synthetic aperture radar (SAR) images detection and classification recently while different kinds of adversarial attacks from malicious adversary and the hidden vulnerability of DNNs may lead to serious security threats. The state-of-the-art DNNs-based SAR image detection models are designed manually by only considering the test accuracy performance on clean datasets but neglecting the models' adversarial robustness under various types of adversarial attacks. In order to obtain the best trade-off between the clean accuracy and adversarial robustness in robust convolutional neural networks (CNNs)-based SAR image classification models, this work makes the first attempt to develop a multi-objective adversarially robust CNN, called MoAR-CNN. In the MoAR-CNN, we propose a multi-objective automatic design method of the cells-based neural architectures and some critical hyperparameters such as the optimizer type and learning rate of CNNs. A Squeeze-and-Excitation (SE) layer is introduced after each cell to improve the computational efficiency and robustness. The experiments on FUSAR-Ship and OpenSARShip datasets against seven types of adversarial attacks have been implemented to demonstrate the superiority of the proposed MoAR-CNN to six classical manually designed CNNs and four robust neural architectures search methods in terms of clean accuracy, adversarial accuracy, and model size. Furthermore, we also demonstrate the advantages of using SE layer in MoAR-CNN, the transferability of MoAR-CNN, search costs, adversarial training, and the developed NSGA-II in MoAR-CNN through experiments.