EUVIMEDEUVIMED entdecken
Crossref

GHP: Hardening Graph‐Based Reranking Defense for Poisoned Retrieval‐Augmented Generation

Wei She, Zikai Dong, Yongkang Yang, Zhao Tian, Wei Liu, Defeng Kong

Concurrency and Computation: Practice and Experience · 2026-07 · 38 (13)

DOI: 10.1002/cpe.70824 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Concurrency and Computation: Practice and Experience, online erschienen am 03.07.2026. Verfasst von Wei She und 5 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Retrieval‐augmented generation (RAG) improves factual grounding by conditioning generation on external evidence, but it also makes answer quality more dependent on retrieval robustness. Under poisoned retrieval, adversarial passages may survive ranking and enter the final prompt, creating a reliability risk even when the generator itself is unchanged. Graph‐based reranking is a promising retrieval‐stage defense because it can exploit structural relations among retrieved candidates. However, our analysis shows that this defense line remains vulnerable when poisoned passages form small near‐duplicate groups that are both strongly query‐supported and locally cohesive. To address this residual failure mode, we propose GHP (graph hardening penalty). Built on a locally restored HRSIM graph‐based reranking baseline, GHP operates at a fixed reranking‐stage interface and leaves retrieval and generation unchanged, while demoting candidates that jointly exhibit strong query support and strong local cohesion. We also introduce GHP‐Lite, a selective inspection‐scope variant that narrows the suspiciousness‐inspection scope while preserving the same fixed final top‐5 evidence interface. Across the main attacked‐setting matrix, spanning two backbones, three datasets, and two poisoning attacks under one shared local evaluation family, the restored HRSIM baseline consistently improves over no defense and serves as a credible graph‐based reference line. On the NQ anchor line with Qwen2.5‐7B‐Instruct under LM‐targeted poisoning, HRSIM reduces attack success rate (ASR) from 50.3% to 5.3% and improves exact match (EM) from 26.7% to 41.3%. GHP‐R further reduces ASR to 1.0% while increasing EM to 44.7%, and a pooled query‐level significance check shows that its robustness gain over restored HRSIM is statistically significant on this anchor setting. GHP‐Lite closely preserves the milder comparison point under a narrower inspection scope, while a simple TF‐IDF cosine near‐duplicate suppression baseline remains close to the no‐defense line and therefore does not account for the gains of the graph‐based line. Benign full‐corpus evaluation shows that clean‐side behavior varies substantially across datasets rather than following a uniform non‐degradation pattern, with the largest penalty concentrated on HotpotQA. Supplementary transfer and attack‐diversity checks broaden the empirical picture, but remain supportive rather than conclusive. Overall, the results suggest that GHP is a practical failure‐mode‐driven hardening extension for restored graph‐based poisoned‐RAG defense under the current local setting, while also highlighting the importance of operating‐point selection across different evaluation conditions.

Zitieren

APA

She, W., Dong, Z., Yang, Y., Tian, Z., Liu, W., & Kong, D. (2026). GHP: Hardening Graph‐Based Reranking Defense for Poisoned Retrieval‐Augmented Generation. Concurrency and Computation: Practice and Experience, 38(13). https://doi.org/10.1002/cpe.70824

Vancouver

She W, Dong Z, Yang Y, Tian Z, Liu W, Kong D. GHP: Hardening Graph‐Based Reranking Defense for Poisoned Retrieval‐Augmented Generation. Concurrency and Computation: Practice and Experience. 2026;38(13). doi:10.1002/cpe.70824

BibTeX
@article{She2026,
  author = {She, Wei and Dong, Zikai and Yang, Yongkang and Tian, Zhao and Liu, Wei and Kong, Defeng},
  title = {GHP: Hardening Graph‐Based Reranking Defense for Poisoned Retrieval‐Augmented Generation},
  journal = {Concurrency and Computation: Practice and Experience},
  year = {2026},
  volume = {38},
  number = {13},
  doi = {10.1002/cpe.70824}
}
Kostenlos starten →