FraudLens AI: A Bank Fraud Detection System Combining Graph Analysis and Generative AI

FraudLens AI is a bank fraud detection platform integrating machine learning, graph analysis, and generative AI, focusing on identifying suspicious mule accounts and hidden fraud ring networks. The project is developed and maintained by Kartikchikni, sourced from GitHub (original title: Fraud-Ring-Detection-using-Graph-AI, link: https://github.com/Kartikchikni/Fraud-Ring-Detection-using-Graph-AI), released on 2026-06-11. This system breaks through the limitations of traditional anti-fraud methods, analyzes account relationships from a network perspective, and improves the efficiency and accuracy of fraud detection.

Modern banking fraud has evolved into complex gang network crimes. Traditional rule-based detection systems can only capture single suspicious transactions and struggle to uncover hidden fraud rings. These gangs use a large number of mule accounts to disperse fund flows; individual transactions seem compliant, but collectively form a money laundering chain. Financial institutions need intelligent methods to understand account relationship networks and detect collaborative fraud behaviors, which is exactly where the combination of graph analysis and AI comes into play.

Graph Network Analysis Engine

The underlying high-performance graph database stores transaction relationships, and regularly performs community discovery (identifying tight account groups), centrality analysis (finding key nodes), and abnormal subgraph detection (discovering abnormal structures).

Machine Learning Risk Scoring

Extract hundreds of features such as transaction frequency and amount distribution, use time-series models (LSTM/Transformer) to capture behavioral anomalies, and integrate graph structure and traditional features to output a 0-100 risk score.

Generative AI Report Module

Automatically generate investigation reports, including fraud ring structure visualization, key account role analysis, transaction timeline, and investigation recommendations, reducing manual workload.

Mule account identification combines transaction behavior and network position indicators: Transaction Behavior Indicators: Fast in-and-out of funds, scattered transaction counterparts, abnormal time patterns (non-normal time periods/time zone jumps); Network Position Indicators: Excessively high betweenness centrality, community marginality (crossing multiple communities), imbalance between in-degree and out-degree. The system uses these indicators to mark suspicious mule accounts with high precision and include them in fraud ring correlation analysis.

Value for financial institutions:

• Economic Benefits: Early detection of fraud rings to avoid huge losses;
• Compliance Value: Provide auditable detection logic and evidence chains to meet AML regulations;
• Operational Efficiency: Shorten manual investigations from weeks to hours;
• Customer Experience: Reduce false positives for normal customers, avoid account freezes and complaints.

Core Challenges and Responses:

• Large-scale Graph Computing: May use graph sampling, hierarchical computing, or incremental update optimization;
• Balance Between Real-time and Accuracy: Find a balance between real-time detection and in-depth analysis;
• Interpretability Requirements: The generative AI report module solves the interpretation problem of black-box models.

FraudLens AI represents an important direction in fintech, organically combining graph neural networks, traditional machine learning, and generative AI to build an intelligent risk control system. For developers, it is a reference case for graph analysis applications; for financial institutions, it enhances their ability to combat complex fraud. In the future, innovations such as cross-institutional federated learning to share fraud information and GNN to improve fraud ring detection accuracy can be expected.