Detecting Sentiment Steering Attacks on RAG-enabled Large Language Models

Retrieval Augmented Generation (RAG) is a technique that enhances the accuracy and reliability of Large Language Models (LLMs), enabling them to answer questions about data they weren't trained on by fetching relevant documents and adding them as context to the prompts sent to an LLM. As companies are incorporating RAG in their LLM applications at a large scale, the technique opens a new attack surface, and sentiment-steering data poisoning attacks are becoming one of the most critical vulnerabilities in RAG, where an attacker can inject biased data to influence the sentiment of an LLM towards an open-ended topic. However, there is a lack of research detecting sentiment-steering attacks. This work lays a foundation towards this by developing a novel detection model to identify poisoned biased passages before they can make their way into databases. From the results, we observe that the detection model achieves a high recall score of 97.75%.