Although Ethereum has shown impressive innovation in blockchain technology, its open and anonymous nature also provides opportunities for malicious actors. The data is clear: by the first half of last year, the total losses from ICO fraud and online scams reached $225 million, and the amounts involved in those abnormal transactions throughout 2022 are even more staggering—$23.8 billion.

How to fill such a huge black hole? Traditional manual annotation methods are no longer sufficient; data labeling is limited, costly, and difficult to scale. Fortunately, breakthroughs in unsupervised machine learning technology in recent years have brought new ideas to transaction risk detection.

**Where to start? First, organize the data**
The first step in abnormal transaction detection is to structure the massive transaction data of Ethereum. Pull historical transaction records from on-chain data sources, then construct a weighted multi-transaction network. Simply put, the nodes in the graph represent accounts, edges represent transaction flows, and weights are determined by transaction amounts and timestamps. The advantage of this design is that it can reveal both the relationships between accounts and the spatiotemporal features of transactions. For example, high-frequency large transactions may indicate a Ponzi scheme, while dispersed small transactions could involve dust attacks.

**Then, use dual graph transformation to extract deep features**
In the original transaction network, transaction relationships are often obscure. Researchers came up with the idea of dual graph transformation: treat each transaction itself as an independent node and establish connections between transactions sharing accounts. With this transformation, the indirect relationships between transactions become explicit, allowing the discovery of many previously unseen chains of relationships.