Backdoor Attacks and Defenses on Large Multimodal Models: A Survey

Recent advances in Large Multimodal Models (LMMs) have achieved remarkable success across diverse downstream tasks. However, recent studies show that these models are highly vulnerable to backdoor attacks. Attackers may implant activatable triggers into the models to induce attacker-specified outputs. Compared with unimodal models, the multimodal nature of LMMs introduces new attack and defense challenges. To provide a landscape of such threats, we conduct a systematic survey of backdoor attacks and defenses in LMMs. Our study covers four major types of LMMs: Vision-Language Pretrained Models (VLPs), Text-Conditioned Diffusion Models (TDMs), Large Vision Language Models (LVLMs), and VLM-based Embodied AI. From the attack perspective, we organize existing studies based on their technical characteristics, which mainly include data poisoning, loss manipulation, and model editing. From the defense perspective, we categorize defense methods along the model development lifecycle, comprising the data preprocessing phase, training phase, and post-hoc phase. Finally, we highlight key trends and open challenges in current research. We hope this survey can serve as a foundation for promoting backdoor studies and developing more reliable LMMs in real-world deployments. A curated list of related resources is available at https://github.com/Robin-WZQ/Awesome-Backdoor-on-LMMs.