Object detectors have many applications in safety-critical systems, but they are sensitive to worst-case perturbations such as adversarial attacks, limiting their real-world applicability. Compared to classification, adversarial robustness in object detection has received less attention, and existing methods often rely on adversarial training, whose performance may not transfer across attacks, perturbation budgets, or architectures.
In this work, we introduce Lipschitz-constrained variants of object detection architectures as robust-by-design alternatives to standard detectors. We validate this approach with LipSSD, a Lipschitz-constrained Single Shot MultiBox Detector (SSD), and conduct a comprehensive study of its adversarial robustness using multiple white-box adversarial attacks and datasets.
We first analyze the accuracy-robustness trade-off induced by Lipschitz constraints, showing it can be controlled through a single training hyperparameter. We then demonstrate that Lipschitz-constrained detectors complement adversarial training: under the same training setup on the Pascal VOC dataset, adversarially trained LipSSD improves mAP@50 on unseen attacks by up to 15 points over classical adversarially trained SSD.
Finally, we utilize specific safety-critical datasets such as LARD and KITTI, demonstrating that Lipschitz-constrained detectors can enhance robustness while largely preserving clean performance. These results suggest that architectural Lipschitz control is a practical and attack-agnostic direction for improving the robustness of object detectors.
Blogger's Review: LipSSD effectively enhances the adversarial robustness of object detectors through Lipschitz constraints, showcasing its potential in safety-critical applications. This method not only improves robustness but also maintains good accuracy, making it worth further exploration and application in future research.