Directed Graph Neural Networks for Anomaly Detection of Smart Ponzi Schemes

Ponzi schemes remain a significant challenge to financial security, particularly on blockchain platforms such as Ethereum, where the autonomy of smart contracts facilitates fraudulent activities. Existing detection methods, typically framed as binary classification tasks, often face the challenge of extreme class imbalance, while conventional graph-based detection methods fail to capture asymmetric transaction dynamics. To address these limitations, we introduce Directed Graph Neural Networks for Anomaly Detection of Smart Ponzi Schemes (DGAD-SPS), a novel approach that leverages directed graph analysis to detect Ponzi schemes in Ethereum's transactional data. By formulating the problem as an anomaly detection task on a directed graph, DGAD-SPS captures the asymmetrical and directional nature of Ethereum transactions, enabling a more accurate differentiation between fraudulent and legitimate contracts. The proposed model employs a self-supervised learning paradigm that combines contrastive and generative learning to derive node embeddings without relying on labeled data, making it particularly well-suited for imbalanced datasets. Experimental validation confirms DGAD-SPS's effectiveness in real-world Ponzi scheme detection through explicit modeling of directional transaction relationships and robust performance under severe data imbalance conditions.