Cross-Domain Fraud Detection in Ethereum Tokens Using On-Chain Time Series Based on Neural Networks and Reservoir Computing

Abstract

Fraud in Ethereum token launches (ICOs, IDOs, and NFTs) produces observable on-chain behavioral patterns: anomalous asset concentration, artificial participant influx, fee manipulation, and abrupt activity collapse. This paper investigates whether four time series used for ICO fraud detection act as transferable behavioral indicators across launch modalities. The main contribution is the cross-domain evaluation protocol itself, applied uniformly to all model families: we empirically evaluate it on 758 labeled assets (ICO, IDO, NFT), comparing Keras baselines (MLP, CNN-MLP, LSTM-MLP), classical Echo State Networks (ESN), and Quantum Reservoir Computing (QRC) under a unified fixed decision threshold (τ = 0.5). Performance is assessed with metrics robust to class imbalance (PR-AUC, ROC-AUC, F1, precision, recall) and a threshold-collapse diagnostic. IDO time series were fully extracted via Etherscan API v2; NFT reached 87.6% extraction coverage (219 of 250 labeled collections). Results show strong intra-domain predictivity across all three domains, but zero-shot transfer fails (ICO→IDO/NFT PR-AUC≈ 0.51, below the ≈ 0.60 random baseline expected at observed prevalence): behavioral invariance was not observed across the domains and fraud definitions considered in this study. ESN-fusion-matched achieved the best ICO discriminative ranking (PR-AUC= 0.808), with no significant difference between classical and quantum reservoirs under the paired statistical protocol.