Improving the Adversarial Robustness of Object Detection with Contrastive Learning

Object detection plays a crucial role and has wide-ranging applications in computer vision. Nevertheless, object detectors are susceptible to adversarial examples. Some works have been presented to improve the adversarial robustness of object detectors, which, however, often come at the loss of some prediction accuracy. In this paper, we propose a novel adversarial training method that integrates the contrastive learning into the training process to reduce the loss of accuracy. Specifically, we add a contrastive learning module to the primary feature extraction backbone of the target object detector to extract contrastive features. During the training process, the contrastive loss and detection loss are used together to guide the training of detectors. Contrastive learning ensures that clean and adversarial examples are more clustered and are further away from decision boundaries in the high-level feature space, thus increasing the cost of adversarial examples crossing decision boundaries. Numerous experiments on PASCAL-VOC and MS-COCO have shown that our proposed method achieves significantly superior defense performance.