Evolutionary Competitive Multiobjective Multitasking Adversarial Scatterer Generation for Attacking SAR Automatic Target Recognition
提出一种进化竞争性多目标多任务方法，同时优化不同数量的SAR散射体参数，在攻击深度学习的SAR自动目标识别时平衡攻击效果和隐蔽性，实验表明能生成高攻击性能且物理可解释的对抗样本。
Adversarial scatterer generation manipulates the parameters of synthetic aperture radar (SAR) scatterers to attack the deep learning-based SAR automatic target recognition (ATR). However, it is difficult to concurrently optimize these parameters of scatterers to balance attack effectiveness and concealment with different numbers of scatterers in a black-box adversarial attack scenario. This paper proposes an evolutionary competitive multiobjective multitasking adversarial scatterer generation (MTASG) method to optimize the physical parameters of a varying number of adversarial scatterers for attacking the deep learning-based SAR ATR. In MTASG, multiple adversarial scatterer generation tasks with different numbers of scatterers are optimized simultaneously in a competitive manner and each task aims to concurrently improve the attack effectiveness and reduce the magnitude of the perturbations. Since all tasks share an identical objective space, a global Pareto front is obtained by identifying the heterogeneous Pareto fronts of all tasks with inter-task competition. To further speed up the convergence, the group-based crossover operators with adaptive selection are designed to dynamically choose the most effective crossover operator. Experimental results on the MSTAR dataset demonstrate that MTASG is able to generate adversarial samples with high attack performance and inherent physical interpretability.