Global confidence degree based graph neural network for financial fraud detectionGlobal confidence degree based graph neural network for financial fraud detection

Date: April 24, 2025

Researchers at Tongji University have unveiled a novel framework, the Global Confidence Degree-based Graph Neural Network (GCD-GNN), designed to significantly improve the detection of financial fraud in complex transaction networks. Published in Artificial Intelligence and Autonomous Systems (AIAS), this innovative approach addresses critical challenges in identifying fraudulent activities, such as camouflaged transactions and intricate relational patterns, by integrating global confidence metrics with advanced graph learning techniques.

Financial fraud remains a pervasive issue, costing enterprises an estimated 3% of their value annually and impacting individuals through sophisticated scams. Traditional graph neural networks (GNNs) often struggle to detect fraud due to two key challenges:

1. ​Complex Relationships: Fraudulent transactions are often hidden within dense, multi-relational networks.
2. ​Camouflage Tactics: Fraudsters deliberately mimic legitimate behavior to evade detection.

Existing GNN-based methods focus on local node interactions but fail to account for global patterns or the "typicality" of nodes within the broader graph context. The GCD-GNN framework introduces a transformative solution by evaluating each node’s ​Global Confidence Degree (GCD)​—a measure of its similarity to a learned prototype representing normal or fraudulent behavior.

The GCD-GNN framework revolutionizes financial fraud detection through three interconnected mechanisms designed to address the limitations of traditional graph-based methods. Central to its innovation is ​prototype learning, which transforms raw node features into global representations that encapsulate the defining characteristics of benign and fraudulent behaviors. These prototypes serve as reference points for evaluating each node’s ​Global Confidence Degree (GCD)​, a novel metric that quantifies how closely a node aligns with its designated prototype. High GCD values indicate typical, trustworthy patterns, while low GCD values signal atypical or suspicious activity. Building on this foundation, the framework employs ​dual-perspective aggregation, strategically combining insights from both high-GCD (typical) and low-GCD (atypical) nodes. This approach not only captures nuanced behavioral signals but also filters out noise, enabling the model to detect sophisticated camouflage tactics. For resource-constrained environments, the lightweight variant ​GCD-GNNlight​ optimizes computational efficiency by streamlining prototype learning and aggregation processes, achieving near-instant inference speeds without sacrificing accuracy.

The framework’s effectiveness was rigorously validated across real-world financial datasets. On ​T-Finance, a benchmark for transaction fraud detection, GCD-GNN achieved a record-breaking 97.26% AUC, outperforming state-of-the-art models like PMP (97.07%) and GHRN (95.78%). Its prowess extended to ​FDCompCN, a dataset tracking corporate fraud in Chinese markets, where it attained 71.72% AUC—surpassing baseline models by significant margins. Beyond raw detection power, GCD-GNNlight demonstrated remarkable operational efficiency, slashing training times by 60% compared to competitors while maintaining robust accuracy.

For further details about the GCD-GNN framework and its implications for autonomous vehicle technology, please refer to the full article published in AIAS: Read the Article.

Journal: Artificial Intelligence and Autonomous Systems (AIAS)
DOI: 10.55092/aias20250004
Method of Research: Experimental study
Subject of Research: Technology
Article Title: Global Confidence Degree Based Graph Neural Network for Financial Fraud Detection
Article Publication Date: 23-Apr-2024