Object detection is a critical component of various security-sensitive applications, such as autonomous driving and video surveillance. However, existing deep learning-based object detectors are vulnerable to adversarial attacks, which poses a significant challenge to their reliability and safety. Through experiments, we found that existing works on improving the adversarial robustness of object detectors have given a false sense of security. We argue that using adversarially pre-trained backbone networks is essential for enhancing the adversarial robustness of object detectors. We propose a simple yet effective recipe for fast adversarial fine-tuning on object detectors with adversarially pre-trained backbones. Without any modifications to the structure of object detectors, our recipe achieved significantly better adversarial robustness than previous works. Moreover, we explore the potential of different modern object detectors to improve adversarial robustness using our recipe and demonstrate several interesting findings. Our empirical results set a new milestone and deepen the understanding of adversarially robust object detection. Code and trained checkpoints will be publicly available.

翻译：目标检测是各类安全敏感应用（如自动驾驶和视频监控）的关键组成部分。然而，现有的基于深度学习的目标检测器容易受到对抗性攻击，这对其可靠性和安全性构成了重大挑战。通过实验，我们发现现有提升目标检测器对抗鲁棒性的工作产生了一种虚假的安全感。我们提出，使用经过对抗预训练的骨干网络对于增强目标检测器的对抗鲁棒性至关重要。我们提出了一种简单而有效的方案，用于对配备对抗预训练骨干网络的目标检测器进行快速对抗微调。在不修改目标检测器结构的情况下，我们的方案实现了显著优于以往工作的对抗鲁棒性。此外，我们探究了不同现代目标检测器利用本方案提升对抗鲁棒性的潜力，并展示了若干有趣发现。我们的实证结果树立了新的里程碑，并深化了对对抗鲁棒目标检测的理解。相关代码和训练好的检查点将公开提供。