Lipschitz-Constrained Detectors Boost Object Detection Robustness Against Attacks.

Object detection systems are crucial for many applications, especially in safety-critical domains, but they are vulnerable to adversarial attacks. Current methods often rely on adversarial training, which may not generalize well across different attacks or architectures. This paper proposes a novel approach using Lipschitz-constrained variants of standard object detectors, making them robust by design. The researchers developed LipSSD, a Lipschitz-constrained version of the SSD architecture, and thoroughly evaluated its adversarial robustness against various white-box attacks and datasets. They found that Lipschitz constraints offer a controllable trade-off between accuracy and robustness, managed by a single hyperparameter. Notably, LipSSD complements adversarial training, showing significant improvements in mean Average Precision (mAP@50) on unseen attacks when combined. It also maintains high clean performance on specialized safety-critical datasets like LARD and KITTI, suggesting that architectural Lipschitz control is a practical and attack-agnostic method for enhancing object detector resilience.