Stealth Backdoor Attack for Remote Sensing Image Classification

Deep neural network models in remote sensing image classification tasks face the threat of backdoor attacks. The models attacked by backdoors can identify the injected trigger sample as the target label designed by the adversary, while the classification results for normal samples (samples without triggers) will not be affected. This paper proposes a stealth backdoor attack for remote sensing image classification. Specifically, we apply the rich geological information in remote sensing images to inject triggers that integrate with the image background, which are not easily noticeable. To improve the effectiveness of the attack, we perform targeted adversarial perturbations on the backdoor samples. During the training phase, both poisoned and clean samples are utilized as data for training the classification model. In the inference phase, the backdoor samples will be classified as the backdoor target category instead of the original category. Experiments on real-world datasets verify that the stealth backdoor attack method has strong attack effectiveness and stealthiness.