Abstract：

Financial fraud causes substantial economic losses, and traditional detection methods struggle to address increasingly sophisticated fraud patterns. This paper systematically reviews the application of intelligent algorithms in fraud detection, covering high-quality research from 2022 to 2025. Evidence shows that ensemble learning achieves AUC 0.99 on structured data, hybrid deep learning models reach AUC 0.9995, and graph neural networks (GNN) improve anti-money laundering (AML) detection accuracy from 85% to 96%. We identify six major research gaps, including label delay, concept drift, and adversarial attack vulnerability, noting that 87% of current studies use static datasets that fail to address real-time evolving fraud patterns. Looking forward, four frontier directions emerge: large language model (LLM) financial applications, quantum computing, federated learning (FL) at scale, and the fusion of GNN with reinforcement learning (RL). This work provides a method selection guide for researchers and a technical roadmap for practitioners, advancing the development of more intelligent, transparent, and robust anti-fraud systems.

Key Words：

financial fraud detection; machine learning; deep learning; graph neural networks; explainable AI