Robustness of Intelligent Transport Systems Employing Computer Vision for the Analysis of Railway Infrastructure Images

With the growing integration of AI systems in the railway sector, it is becoming essential to ensure their robustness against external interference, particularly subtle alterations in input data. This paper analyses the resilience of various neural network architectures, including ResNet18, ResNet50,VisionTransformer, a convolutional neural network, and the GPT-4o multimodal model, which are used for automated fault detection in the analysis of railway infrastructure images. Experiments have been conducted using adversarial disturbances generated via universal noise derived from a set of transformers. Two attack modifications were employed to simulate real-world interference scenarios with limited data availability. The performance of the models has been evaluated on both untainted images and those compromised by overlaid noise. The outcomes indicate that while ResNet50attains maximum accuracy on unmodified data, ViT and GPT-4o demonstrate greater resilience to adversarial disturbances. The research emphasizes the importance of selecting model architectures based on both their accuracy and their robustness against distortions. This paper introduces a method for assessing robustness and offers practical recommendations for developing AI systems designed for application in railway environments where safety is critical.