Sponge Backdoor Attack: Increasing the Latency of Object Detection Exploiting Non-Maximum Suppression

Backdoor attacks against deep learning based object detectors have been studied increasingly in recent years. While most proposed attacks primarily focus on compromising the model’s integrity by inducing incorrect detections, only few studies explore backdoor attacks targeting the model’s availability, a critical concern in safety-critical domains such as autonomous driving. In this paper, we introduce a novel backdoor attack called the Sponge Backdoor Attack (SBA), designed to increase the detection latency of end-to-end object detectors. Specifically, we overload a commonly employed technique in many object detectors - non-maximum suppression (NMS) by introducing a large amount of non-existent objects. Through comprehensive experiments, we demonstrate the SBA’s effectiveness to prolong the processing time of the poisoned image while maintaining detection performance on clean images across various models, datasets, and hardware platforms.