Transferable Adversarial Camouflage Attacks with Multi-layer Topology Tuning for Intelligent Vehicles

Deep learning has been widely used in intelligent vehicle perception systems, but its vulnerability to adversarial attacks poses a serious threat to system security. Especially in the physical world, existing adversarial camouflage attacks often face problems such as poor cross-model transferability and visual unnaturalness, which limits their practical feasibility and concealment. To this end, this paper proposes a transferable adversarial camouflage generation method based on multi-layer topology adjustment. Based on the three-dimensional model of the vehicle, this method constructs a multi-layer attention graph and uses depth-first search (DFS) to generate a connectivity graph of the attention region. By adjusting the topological structure of the graph, the perturbation is guided to be reasonably distributed in different attention layers, thereby improving the interference effect on multiple detection models. At the same time, smoothness loss and unprintability loss are introduced to enhance the naturalness and feasibility of camouflage in physical scenes. Extensive experiments in digital and physical test environments show that this method significantly improves the cross-model transfer performance while maintaining the effectiveness of the attack. In the physical attack task, compared with the existing methods, the attack success rates on Faster R-CNN, Mask R-CNN and Deformable DETR reached 56.3%, 48.5% and 93.4% respectively, all achieving the best performance.