Vis-Poison: Poisoning Visual Knowledge in Multimodal Retrieval-Augmented Generation

Abstract

While multimodal retrieval-augmented generation (RAG) systems increasingly rely on images as external knowledge sources, the introduction of poisoned visual evidence can severely compromise multimodal large language model (MLLM) generation. Unlike prior attacks that rely on altering textual metadata, we introduce Vis-Poison, a novel visual knowledge poisoning attack where the poisoned image itself is the attacker-controlled payload, without manipulating captions, summaries, metadata, or other as

Categories

Framework mappings

Suggested from the entry's categories.

@misc{liang2026vispoison,
  title = {{Vis-Poison: Poisoning Visual Knowledge in Multimodal Retrieval-Augmented Generation}},
  author = {Ru-Jin Liang and Zhong-Pu Chen and Yuhao Lei and Xin Miao},
  year = {2026},
  month = aug,
  eprint = {2608.20756},
  archivePrefix = {arXiv},
  url = {https://www.semanticscholar.org/paper/cbe9563fd7efd2df8017dd23c0e4a6e31b9fc5da}
}