Using Graph Neural Networks (GNNs) for Fraud Detection in Decentralized Finance (DeFi) Networks

Abstract

Decentralized finance (DeFi) apps enable the delivery of lending, trading, asset exchange and more, delivered via blockchain based smart contracts. The transparency and composability provide the environment for innovation, but so do the opportunities for fraud, from phishing, rug pulls and wash trading, to coordinated manipulation of funds across addresses and protocols, and more. Traditional detection systems tend to be isolated and consider transaction or address based decisions without taking into account the relationship between suspicious entities. In this paper, a novel graph neural network (GNN) framework is proposed to detect suspicious activities in DeFi networks through a heterogeneous and temporal graph representation of blockchain addresses, smart contracts, tokens, and transactions. Transaction-flow features, temporal encodings, graph attention, and multi-task prediction are all incorporated into the framework, which is designed to provide an estimation of risk at the address, transaction, and protocol level. It includes class-imbalance handling, temporal validation, explainability, uncertainty estimates and analyst review to minimize false-positive and overconfident automated decisions. A research methodology is defined for comparison with rule-based systems, as well as gradient-boosted trees and non-graph based systems. The following measures are used to evaluate: precision, recall, F1-score, area under the precision–recall curve, detection delay, and false-positive rates at set operating thresholds. The paper does not contain any fabricated empirical data but rather a proposed architecture and an evaluation plan. It states that GNNs can serve as a valuable relational representation for DeFi fraud analysis, but they are only effective if they are well-labeled, are resistant to temporal drift and adversarial drift, and are properly governed around alerts and investigations.