A1974
Title: Extreme value-informed multi-modal transfer learning for blockchain fraud detection
Authors: Jeffrey Chu - Renmin University of China (China) [presenting]
Abstract: Blockchain-based financial systems face increasingly sophisticated fraudulent schemes, yet detecting anomalies in large-scale transaction graphs remains challenging due to high dimensionality, non-stationary behaviour, and Multimodal structure. An Extreme-Value-Informed Multimodal Transfer Learning (EVI-MTL) framework is proposed that integrates structural transaction features, participant-level temporal sequences, textual and video metadata, and extreme value-based risk measures. A Peaks-over-Threshold model is used to construct tail-risk embeddings and a temporally-aware Maximum Mean Discrepancy (TMMD) regularization to address distributional shifts across chains and over time. Empirical studies on Ethereum and Solana show that EVI-MTL substantially improves fraudulent-node classification, outperforming state-of-the-art baselines by 5-15 percentage points, with the largest gains during periods of elevated tail risk. Results indicate that incorporating extreme-value characteristics meaningfully enhances Transfer Learning and yields more reliable early-warning indicators for phishing, rug pulls, and related illicit behaviours, providing a statistically grounded approach for monitoring risk-sensitive blockchain environments.
