A multi-dimensional collaborative strategy for Robust lane detection under AI safety consideration

AI safety has become a major barrier to the reliable deployment of autonomous driving systems. Deep learning-based perception models often suffer from performance degradation under complex environments, dynamic conditions, and abnormal inputs. As a key perception task, lane detection is highly sensitive to such challenges, where failures can lead to vehicle deviation, planning errors, and serious safety risks. However, existing defense methods struggle to balance real-time performance and high accuracy for continuous video streams, and most rely on single-dimensional strategies that are ineffective in complex traffic scenarios. To address these limitations, we propose a Multi-Dimensional Collaborative Defense (MCD) mechanism, which builds a three-dimensional defense framework integrating data, model, and training strategies to comprehensively enhance the robustness and safety of lane detection systems. At the data dimension, a lane feature extractor fusing gradient sensitivity and channel importance generates adaptive weight coefficients for block-wise input purification. The model dimension introduces local gradient-guided adversarial samples to optimize attack feature recognition. During training, a scene weight allocation mechanism enhances adaptability to diverse lighting and road conditions. Experiments on the CULane dataset show MCD improves detection accuracy by 27.3% on average versus undefended models. Under extreme conditions (e.g., crowded traffic, low light), its F-score outperforms single defense methods by up to 34%, verifying its dual enhancement of attack robustness and environmental generalization.