In recent years, phishing scams have emerged as one of the most serious crimes on Ethereum. Existing phishing scam detection methods typically model public transaction records on the blockchain as a graph, and then identify phishing addresses through manual feature extraction or graph learning frameworks. Meanwhile, these methods model transactions within a period as a static network for analysis. Therefore, these methods lack the ability to capture fine-grained time dynamics, and on the other hand, they cannot handle the large-scale and continuously growing transaction data on the Ethereum blockchain, resulting in lower scalability and efficiency. In this paper, we propose a two-dimensional streaming framework 2DynEthNet for Ethereum phishing scam detection. First, we cast the transaction series into 6 slices according to block numbers, treating each as a separate task. In the first dimension, we treat transaction features as edge features instead of node features within one task, allowing each transaction to be streamed in 2DynEthNet, aiming to capture the evolutionary features of the Ethereum transaction network at a fine-grained level in continuous time. In the second dimension, we adopt the strategy of incremental information training between tasks, which utilizes meta-learning to quickly update the model parameters under new slices, thus effectively improving the scalability of the model. Finally, experimental results on large-scale real Ethereum phishing scam datasets show that our 2DynEthNet outperforms the state-of-the-art methods with 28.44% average Recall and achieves the most efficient training speed, proving the effectiveness of both temporal edge representation and meta-learning. In addition, we provide an Ethereum large-scale dynamic graph transaction dataset, ETGraph, which aligns with the data distribution in real transaction scenarios without sampling and filtering unlabeled accounts.

中文翻译：

---

2DynEthNet：用于以太坊网络钓鱼诈骗检测的二维流框架


近年来，网络钓鱼诈骗已成为以太坊上最严重的犯罪之一。现有的网络钓鱼诈骗检测方法通常将区块链上的公共交易记录建模为图形，然后通过手动特征提取或图形学习框架识别网络钓鱼地址。同时，这些方法将一段时间内的交易建模为静态网络进行分析。因此，这些方法缺乏捕捉细粒度时间动态的能力，另一方面，它们无法处理以太坊区块链上大规模且不断增长的交易数据，从而导致可扩展性和效率较低。在本文中，我们提出了一个二维流框架 2DynEthNet 用于以太坊网络钓鱼诈骗检测。首先，我们将交易序列根据区块号转换为 6 个切片，将每个切片视为一个单独的任务。在第一个维度中，我们将交易特征视为一个任务中的边缘特征，而不是节点特征，允许每笔交易在 2DynEthNet 中流式传输，旨在连续时间在细粒度层面捕捉以太坊交易网络的进化特征。在第二个维度上，我们采用了任务间增量信息训练的策略，利用元学习快速更新新切片下的模型参数，从而有效提高模型的可扩展性。最后，在大规模真实以太坊网络钓鱼诈骗数据集上的实验结果表明，我们的 2DynEthNet 以 28.44% 的平均召回率优于最先进的方法，并实现了最有效的训练速度，证明了时间边缘表示和元学习的有效性。 此外，我们还提供了一个以太坊大规模动态图交易数据集 ETGraph，它与真实交易场景中的数据分布保持一致，无需采样和过滤未标记的账户。