This is the official source code for the paper:
"LABRA: A Label-free Graph Neural Framework for Illicit Account Detection in Large-scale Blockchain Transaction Networks" 🔗
Important
🚧 Work in Progress: We are actively updating the source code and datasets to ensure full reproducibility. Stay tuned for more updates! 📦✨
The following figure illustrates the overall architecture of LABRA. Our framework systematically processes blockchain transaction data through a zero-shot learning paradigm to identify illicit accounts.
Figure 1: The overall pipeline of LABRA, showcasing the flow from transaction networks to illicit account detection.
| Category | Method | Ethereum-S (AUROC) | Ethereum-S (AUPRC) | Bitcoin-M (AUROC) | Bitcoin-M (AUPRC) |
|---|---|---|---|---|---|
| Supervised | GCN | 57.03 | 53.96 | 54.55 | 20.71 |
| GAT | 66.19 | 61.93 | 64.44 | 25.10 | |
| SAGE | 66.38 | 59.28 | 63.40 | 21.13 | |
| Semi/Un-sup | GGAD | 64.56 | 63.70 | 56.76 | 24.60 |
| GADAM | 56.85 | 53.55 | 44.19 | 17.06 | |
| FreeGAD | 66.00 | 59.28 | 64.74 | 22.99 | |
| 🚀 Ours | LABRA | 75.20 | 74.65 | 65.09 | 27.87 |
| Improvement | +8.82↑ | +10.95↑ | +0.35↑ | +2.77↑ |