Improving Adversarial Detection Methods for SAR Image via Joint Contrastive Cross-entropy Training
Zhiwei Zhang, Shuowei Liu, Xunzhang Gao, Yujia Diao · 2022
Adversarial examples (AEs) have become a critical security concern for intelligent synthetic aperture radar (SAR) target recognition system. Current defense methods design certain distance metric to characterize the difference between AEs and natural samples in feature space, but expose severe performance degradation against SAR AEs with small perturbation scale. By exploiting Siamesed sample augmentation and feature wised contrastive regularization, we propose a joint contrastive cross-entropy training method to endow natural SAR test samples with more consistent distribution to training samples and higher separability from AEs. Experiments demonstrate that our method significantly improves the performance of current adversarial detections against SAR AEs especially when the perturbation scale is small.