Research
Detecting Poisoning Attacks in the RAG Systems Using Multi-Tier Anomaly Detection
Publication date not yet evaluated · Added sdstate.edu
Page published
Coverage timeline
Single-source research — one report is available.
Why it matters
RAG poisoning attacks inject malicious content into knowledge bases to manipulate LLM outputs, and retrieval-stage detection methods give defenders a way to catch these attacks before they influence generation.
A South Dakota State University master's thesis by Somtochukwu Orizu proposes a multi-tier anomaly detection approach to detect poisoning attacks against Retrieval-Augmented Generation (RAG) systems, reproducing known attacks such as PoisonedRAG and PoisonCraft and building datasets for retrieval-stage detection.