APBAM: Adversarial perturbation-driven backdoor attack in multimodal learning

Shaobo Zhang, Wenli Chen, Xiong Li, Qin Liu, Guojun Wang

Abstract: Highlights•We developed a novel multimodal backdoor APBAM by simultaneously attacking bimodal.•We optimized visual triggers via adversarial perturbation and proposed a text trigger method for double mask modeling.•We conducted a comprehensive comparative analysis of APBAM against other well-known backdoor attack methods.•We illustrated the effectiveness and robustness of the innovative APBAM.

External IDs:dblp:journals/isci/ZhangCLLW25