OOBA: Object Offset Backdoor Attack for Remote Sensing Object Detection

The growing adoption of consumer electronic Unmanned Aerial Vehicle (UAV) systems in outdoor military and civilian scenarios is inseparable from the rapid development of remote sensing object detection (RSOD) technologies. However, the security of RSOD has not yet received sufficient attention, particularly with respect to backdoor attacks. In such attacks, the backdoor model behaves normally on benign data but can be maliciously manipulated using pre-defined triggers, posing a serious threat to the reliable detection of UAV systems. Nevertheless, existing backdoor methods designed for natural images fail to adequately capture the characteristics of remote sensing imagery, and their single data modification mode limits their effectiveness on RSOD tasks. In this paper, we explore mislocalization attacks and propose the first RSOD-specific backdoor method named Object Offset Backdoor Attack (OOBA). The proposed OOBA embeds the backdoor through data poisoning and has two data modification modes: Position Offset and Class Modification. Specifically, Position Offset leverages redundant background content in remote sensing images by offsetting the bounding boxes to induce mislocalization by the detector. Meanwhile, this additional background information contained in the bounding boxes can assist the detector in fully embedding backdoors on tiny objects. Class Modification aims to induce misclassification by the detector, overcoming limitations of single data modification mode. Extensive experiments on the NWPU VHR-10 and DIOR datasets demonstrate the state-of-the-art attack performance of OOBA.