Detecting adaptive wash trading behaviors in NFT marketplaces: a neuro-symbolic approach

Abstract

This study proposes a neuro-symbolic framework to detect increasingly sophisticated NFT wash trading that traditional rule-based systems and standard GNNs struggle to identify. The proposed ResidualGATv2 model combines heuristic logic with spatio-temporal graph learning and incorporates transaction timing through edge-level temporal encoding. Tested on 108,588 real NFT transactions, the model achieved 0.99 ROC-AUC, 0.93 PR-AUC, and 0.87 F1-score, while forensic review confirmed 93.6% of identified suspicious wallets as genuine wash traders. The model also demonstrated strong generalization to previously unseen NFT collections, highlighting its potential for effective blockchain fraud detection.