BLTOO-MFFL: Automated and Adversarially Robust Deep Learning for SAR Image Classification by Bi-Level Three-Objective Optimization and Multi-Feature Fusion Loss

The deep learning-based synthetic aperture radar (SAR) image classification model is vulnerable to adversarial attacks. Although automated deep learning (ADL) including neural architecture search (NAS) and hyperparameter optimization (HPO) can help obtain models with robustness against adversarial examples in natural images, the characteristics of SAR images, such as occlusion, speckle noise, data distribution, and high resolution, make the application of ADL in the SAR field rare. This work proposes an automated design method of an adversarially robust and lightweight deep neural network model termed as BLTOO-MFFL for SAR image classification based on bi-level three-objective optimization and multi-feature fusion loss (MFFL). The model comprises a shallow network with shallow feature losses and two deep networks with deep feature losses, and these networks are stacked by the cell-based neural architectures to be optimized by NAS. The multiple losses from different network depths are weighted and fused to construct a MFFL that guides parameter optimization during the model training. The upper-level optimization is the joint NAS and HPO process, where the cell-based neural architectures, the loss weight parameters used in MFFL, and the learning rate are encoded as the decision variables, and the clean accuracy, the adversarial accuracy, and the number of model parameters are evaluated as the objective functions. Non-dominated Sorting Genetic Algorithm-II (NSGA-II) with a hybrid encoding strategy is developed for three-objective optimization. The lower-level optimization is the training process of the model optimized by the upper-level optimization to minimize the MFFL. Experimental results on the FUSAR-Ship dataset and the OpenSARShip dataset with multi-type adversarial attacks demonstrate the superiority of BLTOO-MFFL to the state-of-the-art methods, such as Darts architecture-based network, NSGA-Net, RobNet, TAM-NAS and MoAR-CNN.