Dual-Branch Sparse Self-Learning With Instance Binding Augmentation for Adversarial Detection in Remote Sensing Images

Remote sensing image analysis technology based on neural networks has significantly facilitated human life. However, adversarial attacks can drastically impair the performance of these models, posing substantial economic and security risks. Current adversarial example (AE) detectors primarily focus on studying attacked natural images, while AEs in the remote sensing domain have not received adequate attention. To address this challenge, we propose a novel dual-branch sparse self-learning framework, leveraging instance binding augmentation. The contrastive branch concurrently enhances intra-instance and inter-example feature discrimination, while the masked branch reconstructs perturbation distributions. Furthermore, our method utilizes sparse encoding within depthwise separable convolutions to efficiently transfer parameters, thereby ensuring compatibility with deployment on mobile devices. Extensive experiments demonstrate that our method achieves state-of-the-art performance in detecting both white-box and black-box attacks on remote sensing images. Specifically, our method achieves an average detection accuracy of 95.69%/95.18%, a recall of 91.8%/93.6%, and an
$F1$
-score of 93.48%/94.22% on two attacked models across various attack scenarios, outperforming existing methods.