Enhancing Fraud Detection via On-Chain Ethereum and Off-Chain X Data Fusion

Abstract

Although Ethereum stands as the dominant blockchain for smart contracts and decentralized applications, faces persistent security challenges from fraudulent activities. Such activities often correlate with off-chain platform, such as blog platform and social media. Existing methods analyze fraud activities primarily rely on on-chain transaction data, neglecting interdependencies between on-chain and off-chain activities. In this paper, we observe that there are associations between airdrop campaigns in X platform, a famous social platform and Ethereum fraudulent activities. Further, we crawl Ethereum addresses and posts of these users in airdrop campaigns, and construct a cross-platform datasets from X to Ethereum, including matching pairs of Ethereum addresses to X users, Ethereum transactions and X post data. Due to inherent heterogeneity between Ethereum transactions (structured graphs) and X data (unstructured text/images), we design a multimodal fusion framework leveraging transformer architectures to fuse on-chain transaction features with off-chain content features (text and image representations). Finally, the fused features are leveraged to construct downstream fraud transactions classifiers. Experimental results demonstrate that classifiers using fused features outperform classifiers using transaction features, achieving a 12% improvement in Recall. Our findings highlight the critical role of off-chain data in enhancing fraud detection accuracy.