TRACE: Transaction representation learning for large-scale Ethereum phishing detection under limited labels

초록

Ethereum phishing detection is a decision-support problem: blockchain risk analysts must prioritize suspicious addresses under severe label scarcity and bounded review capacity. We present TRACE, a two-stage decision-support artifact that constructs dense Ethereum address graphs from Etherscan transaction records, pre-trains an edge-aware encoder with graph contrastive learning, and adapts it with validation-based threshold calibration and explanation analysis for reviewable phishing screening. Rather than proposing new learning primitives, the contribution is a domain-grounded configuration whose individual design choices are non-obvious for dense, heavy-tailed transaction graphs and whose benefit grows with graph scale. Across five graphs from 30,000 to 149,930 addresses, six label ratios, eight baselines, and five random seeds, TRACE achieves the best overall mean performance (F1 0.5094; AUC 0.9365). Component ablations, paired significance tests, precision-recall (AUPRC) reporting, and prevalence-sensitivity experiments down to sub-1% phishing rates confirm that each design element contributes and that the advantage is stable under realistic class imbalance. Performance improves with graph scale, reaching F1 0.6720 and AUC 0.9730 on D150k under full labels, and a smaller pre-training graph transfers to a larger target (F1 0.5670). The artifact logic generalizes to other scarce-label, interaction-rich risk-screening settings beyond Ethereum.

키워드