Fine-grained hierarchical multi-round iterative semantic optimization attack method for RAG systems

Area of Science:

• Artificial Intelligence
• Computer Security
• Natural Language Processing

Background:

• Retrieval-augmented generation (RAG) systems combine information retrieval with large language models (LLMs).
• Knowledge corruption attacks inject malicious text into RAG systems to manipulate LLM outputs.
• Existing attacks suffer from limited concealment due to single-generation strategies and semantic deviation.

Purpose of the Study:

• To propose a novel, fine-grained, hierarchical, multi-round iterative semantic optimization (FHM-ISO) attack method for RAG systems.
• To enhance the effectiveness and concealment of knowledge corruption attacks.
• To overcome limitations of conventional semantic similarity metrics in attack optimization.

Main Methods:

• Developed a multi-round iterative semantic optimization approach (FHM-ISO).
• Utilized dynamic prompt selection based on semantic relationships between questions and generated content.
• Introduced a hybrid similarity scaling mechanism integrating non-linear activation and linear scaling for semantic distance evaluation.

Main Results:

• The FHM-ISO method significantly improves attack effectiveness against RAG systems.
• Achieved an average attack success rate of 68% with single-piece malicious text injection per question.
• Demonstrated superior performance across multiple public datasets compared to traditional methods.

Conclusions:

• The proposed FHM-ISO attack is highly effective in corrupting knowledge within RAG systems.
• The hybrid similarity scaling mechanism enhances semantic optimization for attack generation.
• This research highlights vulnerabilities in RAG systems and proposes a sophisticated attack vector.