Adversarial Attacks for Semantic Segmentation of Autonomous Systems in Unstructured Off-Road Environments

In off-road environments, semantic segmentation models face severe robustness challenges under adverse conditions such as muddy terrain and uneven surfaces. To address these issues, we propose a novel multi-dimensional adversarial attack framework that integrates channel attention, regional structural perception, and edge-guided mechanisms. This method is designed to uncover the inherent vulnerabilities of existing segmentation models from multiple perspectives. To enable more realistic and comprehensive evaluation, we construct a dedicated robustness benchmark dataset for off-road semantic segmentation, which includes complex background distractions and real-world obstacle information. Extensive experiments on two representative segmentation networks demonstrate that our method can substantially degrade model performance while preserving perturbation sparsity and imperceptibility. This study provides a new perspective on evaluating the robustness of deep vision models in challenging off-road scenarios and establishes a foundation for the development of future defense strategies.