From Data to Decision: An Intelligent Framework for Identifying Ponzi Schemes in Decentralized Finance (DeFi) Platforms

The booming development of decentralized finance, while driving financial innovation, has also spawned new forms of fraud such as smart Ponzi schemes. Existing detection methods exhibit significant deficiencies in feature representation, imbalanced sample handling, and early warning capabilities. This paper proposes MM-GAT, a graph attention network detection framework that fuses multimodal features and performs joint modeling from two dimensions: static code semantics and dynamic transaction behavior of smart contracts. The code feature extraction module employs multi-channel convolutional neural networks to capture local patterns in opcode sequences, combined with a Transformer encoder to model long-range dependencies. The transaction feature extraction module constructs an account transaction graph and utilizes graph attention networks for structural encoding. A gated fusion mechanism adaptively adjusts the contribution weights of different modal features according to sample characteristics. The contrastive learning strategy enhances minority class feature representation by maximizing positive sample pair similarity. Experimental results on the Ethereum smart contract dataset show that MM-GAT achieves an F1 score of 94.67% and an AUC value of 97.23%, representing improvements of approximately 11.2 and 8.5 percentage points respectively compared to the best baseline method. Ablation experiments validate the effectiveness of each module, and the model demonstrates robust identification capabilities under extremely imbalanced conditions and early detection scenarios. This research provides a new technical pathway for blockchain financial fraud detection.