FraudShield: Knowledge Graph-Enhanced LLM Defense Against Fraud Attacks

FraudShield constructs and refines a fraud technique-keyword knowledge graph to structure high-confidence associations between suspicious text and fraud techniques, and augments the original input with highlighted keywords and evidence to guide the LLM toward safer responses. Experiments show that this method consistently outperforms existing state-of-the-art defenses across four mainstream LLMs and five representative types of fraud, and provides interpretable clues for model generation.