Attention-Aware Temporal Adversarial Shadows on Traffic Sign Sequences
Published in CVPR 2025 – IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025
We present a framework for black-box adversarial attacks on traffic signs using temporally coherent shadows. Unlike previous methods that target isolated frames, our approach attacks entire traffic sign sequences, mimicking real-world autonomous vehicle (AV) observation behavior.
Our method uses a non-differentiable shadow generator with fixed geometry and opacity, whose spatial scale evolves over time. A genetic algorithm is employed to optimize shadow parameters under a dual-loss objective that maximizes both classification error and attention disruption, using DINO ViT attention maps.
Evaluated on GTSRB, our method achieves a sequence-level attack success rate (SL-ASR) of up to 87.5%, and attention supervision improves attack effectiveness by 11–18% compared to baseline strategies.