MTAttack: a multi-target backdoor attack against large vision-language models

Researchers propose MTAttack, a multi-target backdoor attack framework against large vision-language models (LVLMs), achieving high attack success rates on mainstream benchmarks and clearly outperforming existing attack methods. The authors call it the first multi-target backdoor attack framework, implanting multiple independent triggers corresponding to different attack targets simultaneously in a single round of training. To address the problem of severe feature interference among triggers causing incorrect trigger-to-target mapping, MTAttack jointly optimizes multiple triggers in latent space through two constraints, Proxy Space Partitioning and Trigger Prototype Anchoring, so that each trigger independently maps clean images to a unique proxy class and ensures they are separable from one another. The paper reports that the attack generalizes strongly across datasets and resists existing backdoor defense strategies; the authors believe the results highlight the vulnerability of LVLMs to multi-target backdoor attacks, and the code is open-sourced at https://github.com/mala-lab/MTAttack。