Graph Neural Networks for Blockchain Fraud & Sybil Detection

The Challenge

Blockchain fraud detection relies on brittle, hand-written heuristics that miss coordinated attacks and novel exploit patterns. Labeled datasets are scarce or nonexistent: you can't train a fraud classifier when you don't know which wallets are fraudulent. Yet the transaction graph itself encodes behavioral signals, if you can learn to read them at billion-node scale.

The Approach

Built two end-to-end pipelines from raw blockchain node data: one for Cardano's UTXO model, one for Ethereum's account model. A custom Rust ETL (streaming, O(1) memory) constructs heterogeneous graphs with 4 node types and 23 edge types, including 17 bot-detection features and DeFi primitives. Three self-supervised GNN architectures (BGRL, GraphMAE, HGOT) learn 128-dimensional behavioral embeddings directly from graph structure. No labels required. Wallets that behave similarly cluster together in embedding space, surfacing patterns invisible to rule-based systems.

The Results

The embeddings enable Sybil detection (identifying coordinated bot networks via same-block activity), fraud classification (flagging serial rug-pull deployers and insider dump patterns), entity resolution (clustering wallets controlled by the same actor), and smart-money tracking (scoring traders by realized PnL). The Ethereum pipeline processes the full chain at 3B+ nodes and 15B+ edges; the Rust ETL achieves 60% memory reduction through hashed address storage. Currently applying these capabilities to active detection tasks.