ShipCamou: adversarial camouflage against optical remote sensing image ship detector

Despite the extensive application of Deep Neural Networks (DNNs) across diverse domains, their vulnerability to adversarial examples remains a notable concern. Simultaneously, adversarial patches afford the capability to camouflage one's equipment under DNN-based object detectors. However, the topic has been underexplored in optical remote sensing images (O-RSIs), especially regarding ship imagery. Meanwhile, existing methods for generating adversarial patches exhibit limited attack strategies and fail to effectively adapt to complex conditions during the optimization process. This paper introduces a new type of adaptive patch method for ship physical camouflage (ShipCamou). We designed adaptive components including physical adaptability and target scale adaptability, to meet the adaptability requirements of the adversarial patch under various physical conditions and different ship target scales. Additionally, we developed a new loss function that optimizes the adversarial patch by reducing the average confidence of the target and distorting the detection bounding boxes to deviate from their normal positions, significantly enhancing the attack efficiency of the adversarial patches. Experiments conducted on the HRSC-2016 maritime remote sensing image dataset against a variety of advanced typical object detectors in both white-box and black-box settings have demonstrated the notable adversarial camouflage effect of our proposed ShipCamou under multiple object detectors. Moreover, these effects were observed to transfer across different models. When compared to two typical adversarial patch methods, our approach outperforms them in terms of reducing the average precision and increasing the attack success rate. The results indicate that ShipCamou can deceive object detectors, effectively concealing our ships or other equipment requiring camouflage in O-RSIs.