Reproduce “Joint-GCG: Unified Gradient-Based Poisoning Attacks on Retrieval-Augmented Generation Systems” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haowei Wang, Rupeng Zhang, Junjie Wang, Mingyang Li, Yuekai Huang, Dandan Wang, Qing Wang

Joint-GCG optimizes one poison to enter retrieved context and steer a RAG answer. It aligns retriever and generator gradients through cross-vocabulary projection, gradient tokenization alignment, and adaptive weighted fusion. On NQ with Contriever and Llama3-8B, it achieves 92% generation attack success rate, versus 89% for LIAR and 72% for GCG. Crafting assumes white-box access to both models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.