Abstract
State-of-the-art architectures for automated railway safety surveillance achieve high benchmark accuracies; however, such systems are systematically vulnerable to adversarial perturbations. Current approaches focus on adversarial training, while other aspects, such as architecture design, are overlooked as robustness mechanisms. We propose RobustRailNet, a novel architecture for the robust and high-performance classification of railway obstacles. The proposed approach builds upon domain-relevant characteristics such as high-frequency textures, spatially coherent structures, and global context. RobustRailNet is evaluated against seven white-box adversarial attacks, including gradient-based, optimization-based, and patch-based methods, representing a worst-case assumption for model vulnerability. The evaluation is conducted on a diverse dataset comprising multiple railway obstacle types. RobustRailNet achieves a highly significant mean robustness gain of up to 66.89 percentage points compared to the state-of-the-art while maintaining a mean balanced accuracy of 99.52%. The results demonstrate that adversarial vulnerability arises from architectural design and is not exclusively linked to training data or strategies. Architectural optimization should therefore be seen as a necessity rather than a possibility for the design of robust, safety-critical systems. When developing and deploying such architectures in practice, the influence of the architecture should be taken into account upfront, rather than attempting to achieve robustness exclusively through ex post retraining.
Funding Information
  • Open-Access-Publication-Fund of the Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg

This publication has 47 references indexed in Scilit: