Global-Local Characteristic Excited Cross-Modal Attacks from Images to Videos

The transferability of adversarial examples is the key property in practical black-box scenarios. Currently, numerous methods improve across different models trained on same modality data. investigation generating video with imagebased substitute to attack target models, i.e., cross-modal examples, rarely explored. A few works directly apply image for each frame and no factors especial data are considered, which limits examples. In this paper, we propose an effective method considers both global local characteristics Firstly, from perspective, introduce inter-frame interaction into process induce more diverse stronger gradients rather than perturb separately. Secondly, disrupt inherently correlation frames within a video, prevents model capturing valuable temporal clues. Extensive experiments UCF-101 Kinetics-400 validate proposed significantly improves even surpasses strong baseline using as model. Our source codes available at https://github.com/lwmming/Cross-Modal-Attack.