Enhancing adversarial resilience in semantic caching for secure retrieval augmented generation systems.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning Security

Background:

• Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) offer advanced NLP capabilities but face high computational costs due to repeated query processing.
• Semantic caching mitigates these costs by reusing past responses for similar queries, but existing methods are vulnerable to adversarial attacks that exploit semantic similarity.

Purpose of the Study:

• To investigate security vulnerabilities in current semantic proximity caching systems, exemplified by GPTCache.
• To introduce SAFE-CACHE, a novel, robust semantic caching approach designed to withstand adversarial exploitation.

Main Methods:

• SAFE-CACHE employs a cluster-centroid-based strategy, contrasting with GPTCache's single-query embedding.
• Utilizes unsupervised clustering, statistical noise detection, bi-encoder refinement, and a fine-tuned LLM for intent inference.
• Compares incoming queries to cluster centroids for enhanced semantic validation.

Main Results:

• SAFE-CACHE significantly reduces adversarial attack success rates from 52.77% to 14.27% compared to GPTCache.
• Demonstrates up to a 72% improvement in adversarial resistance.
• Maintains system reliability and security against malicious inputs.

Conclusions:

• Current semantic caching systems like GPTCache are susceptible to adversarial attacks.
• SAFE-CACHE offers a more secure and resilient alternative through its cluster-centroid approach.
• The proposed method effectively mitigates security risks while preserving the benefits of semantic caching for LLMs.