关键词: 限速；交通标识识别；鲁棒性测试；对抗样本；图像变换

Abstract: To efficiently evaluate the robustness of traffic sign recognition algorithms, particularly those for speed limit signs, this paper proposes an adversarial sample generation framework. Traffic signs are processed using various transformation algorithms, including illumination and angle transformations, to simulate environmental changes. The transformed images are embedded into real road backgrounds to mimic realistic traffic scenarios. A deep learning model is employed to process these embedded traffic images containing speed limit signs. Single bounding box and multiple bounding box filters are utilized to screen out bounding boxes for anti-disturbance training. During the training process, four loss functions are designed based on four distinct expected attack effects: target attack, hidden attack, non-target attack, and revealed attack. These functions are used to calculate the loss values. The back-propagation algorithm is then applied to update the adversarial perturbations, ultimately generating adversarial samples. This framework can rapidly generate adversarial samples capable of simulating four types of attacks, thereby assessing the robustness of traffic sign recognition algorithms, especially those for speed limit signs. Furthermore, the generated adversarial examples can be transferred to real-world physical vehicle testing environments, assisting developers in optimizing traffic sign recognition algorithms.

Key words: speed limit; traffic sign recognition; robustness testing; adversarial example; image transformation