ARCADIA : Adaptive Real-time Cross-chain Abuse Detection through Inductive Analysis

Cross-chain systems were once praised for enabling seamless blockchain interoperability; however, they are now exploited for money laundering and hiding digital assets. This reveals a significant gap in decentralized compliance and oversight, primarily resulting from the use of obfuscation techniques. This paper introduces a novel Adaptive Real-time Cross-chain Abuse Detection through Inductive Analysis (ARCADIA) framework, designed to identify exploitation of cross-chain systems for illicit purposes. It uses a hybrid machine learning (ML) architecture that adapts to analyse addresses (with and without history), leveraging temporal features and graph-based inductive reasoning. Our approach enables Virtual Asset Service Providers (VASPs) to deploy ML models for real-time risk label generation. Extensive experiments on Ethereum address datasets have demonstrated superior performance in identifying phishing and malicious entities, even under severe data imbalance conditions. Our framework narrows the gap between blockchain forensics, ML, and compliance, offering a scalable solution aligned with emerging regulatory standards such as the Financial Action Task Force (FATF) Travel Rule.