Intelligent Anti-Fraud System Integrating XAI and Large Language Models: Making AI Decisions Transparent and Explainable

This article introduces an end-to-end anti-fraud system architecture that combines XGBoost prediction, SHAP explainability, and large language model reasoning to address the interpretability dilemma of traditional anti-fraud models, converting technical model decisions into human-understandable explanations to meet multiple needs such as compliance and trust.

In fields like finance and e-commerce, traditional rule engines have strong interpretability but struggle to handle complex fraud methods; black-box models such as deep learning are accurate in prediction but lack explanations. This dilemma is not only a technical issue but also a compliance (e.g., EU GDPR's right to explanation clause, U.S. Fair Credit Reporting Act) and trust issue. Converting complex model decisions into natural language explanations is a huge challenge.

The system adopts a three-in-one architecture:

1. XGBoost Prediction Engine: Processes tabular data. Its performance advantages include being superior to deep neural networks (for tabular data), regularization to prevent overfitting, strong robustness, and fast training. Feature engineering covers transaction, user behavior, relationship, and time-series features.
2. SHAP Explainability Analysis: Provides feature-level attribution, including global (identifying important features) and local (decomposing contributions of individual predictions) explanations, with mathematical rigor.
3. LLM Reasoning Conversion: Converts SHAP technical outputs into natural language explanations. The process is: collect data → build prompts → call LLM to generate reports (e.g., explaining the reasons for transaction anomalies).

Data Flow: Raw transaction data → feature engineering → XGBoost prediction → SHAP explanation → LLM conversion → final report. Prompt Engineering: Needs to include context information (user profile, historical transactions), SHAP results (positive/negative features and their contributions), and output requirements (audience, style, format). Latency Optimization: Asynchronous processing (decoupling decision-making and explanation), caching similar cases, selecting appropriate models (lightweight/powerful models), and streaming output to improve user experience.

1. Risk Control Operation Center: Accelerate alert classification, provide investigation clues, and generate standardized documents.
2. Customer Service Center: Directly cite explanations to respond to customer queries and explain reasons in plain language.
3. Customer Self-Service: Display transparent explanations in the App to reduce complaints and enhance trust.
4. Regulatory Compliance: Show decision explanation chains, fairness analysis, and manual review results.

Hallucination Risk: LLMs may generate inaccurate explanations. Mitigation strategies include limiting reasoning based on SHAP data, requiring prompts not to fabricate information, and fact-checking. Cost Considerations: Large-scale LLM calls are costly. Optimization directions include generating explanations only for complex cases, reusing templates, and locally deploying open-source models. Privacy Protection: Data desensitization, local deployment, and federated learning to avoid exposure of raw data.

Future Directions: Multimodal fraud detection (integrating biometrics, etc.), continuous learning to adapt to new fraud patterns, and enhanced causal reasoning (distinguishing between causation and correlation). Conclusion: This system combines accuracy, transparency, and expressiveness to build a trust bridge between AI and humans. As regulatory and user demands increase, the XAI+LLM architecture will drive AI from a black box to a trustworthy partner.