Reproduce “GRADA: Graph-based Reranking against Adversarial Documents Attack” on OpenResearch

About this paper

Jingjie Zheng, Aryo Pradipta Gema, Giwon Hong, Xuanli He, Pasquale Minervini, Youcheng Sun, Qiongkai Xu

GRADA reranks retrieved documents by comparing them with one another as well as with the question, using a graph and query-aware similarity. In a GPT-3.5-Turbo Natural Questions test with one poisoned document, HRSIM reduced ASR from 55.7±1.2% to 3.0±0.6% and raised EM from 33.3±1.1% to 58.0±1.1%. Results were less stable on multi-hop HotpotQA, and the method assumes poisoned documents are a minority.

Key insights

• Adversarial documents, despite high query relevance, tend to lack semantic coherence with genuinely relevant documents, exhibiting 'semantic isolation' within the retrieved set.
• Graph-based reranking effectively exploits this isolation by allowing benign documents to reinforce each other's scores, forming clusters, while isolating and suppressing adversarial outliers.
• Retrieving a larger initial set of documents and using uniform score initialization for PageRank improves robustness by preventing adversarial documents from gaining an initial advantage.

Summary by alphaXiv.