Lightweight Graph Neural Network for Blockchain Anomaly Detection

Abstract

Blockchain networks operate transactions without any central control which allows attacks like double spending, Sybil attacks, and money laundering serious concerns even with multiple security features. Standard detection methods heavily relies on fixed rules or basic Machine Learning (ML) models that can't adapt well and completely fails to understand the complex transaction relationships. Advanced Graph Neural Network (GNN) models like Graph Convolutional Network (GCN) and Graph Attention Network (GAT) capture these patterns with precision but these models require massive resources like 50-100 GB RAM that ultimately rule out the practical usage on normal systems. This paper combines the GraphSAGE based approach with neighbor sampling in which it samples ten neighbors at the first hop and five at the second, droping noisy edges, and groups transactions into community-aware batches. It will keeps memory under 8 GB. It is built specifically to run on the MacBook Air M1 using PyTorch Geometric's MPS backend, it processes Elliptic dataset subgraphs of roughly 1600 nodes per batch. 93-95% accuracy matched compared to full GNN while using just 385 MB per batch and peak memory never exceed 6.2 GB. XGBoost reach 94-96% accuracy but ignores relational structure of transactions, this catches multi-hop laundering chains traditional methods miss. Exchanges and regulators can now deploy effective fraud detection on standard laptops instead of data center hardware.