Decentralized Applications (DApps) are increasingly developed and deployed on blockchain platforms such as Ethereum. DApp fingerprinting can identify users' visits to specific DApps by analyzing the resulting network traffic, revealing much sensitive information about the users, such as their real identities, financial conditions and religious or political preferences. DApps deployed on the same platform usually adopt the same communication interface and similar traffic encryption settings, making the resulting traffic less discriminative. Existing encrypted traffic classification methods either require hand-crafted and fine-tuning features or suffer from low accuracy. It remains a challenging task to conduct DApp fingerprinting in an accurate and efficient way. In this paper, we present GraphDApp, a novel DApp fingerprinting method using Graph Neural Networks (GNNs). We propose a graph structure named Traffic Interaction Graph (TIG) as an information-rich representation of encrypted DApp flows, which implicitly reserves multiple dimensional features in bidirectional client-server interactions. Using TIG, we turn DApp fingerprinting into a graph classification problem and design a powerful GNN-based classifier. We collect real-world traffic datasets from 1,300 DApps with more than 169,000 flows. The experimental results show that GraphDApp is superior to the other state-of-the-art methods in terms of classification accuracy in both closed- and open-world scenarios. In addition, GraphDApp maintains its high accuracy when being applied to the traditional mobile application classification.

中文翻译：

---

使用图神经网络通过加密流量分析准确识别去中心化应用程序


去中心化应用程序（DApp）越来越多地在以太坊等区块链平台上开发和部署。 DApp 指纹识别可以通过分析产生的网络流量来识别用户对特定 DApp 的访问，从而揭示用户的许多敏感信息，例如他们的真实身份、财务状况以及宗教或政治偏好。部署在同一平台上的DApp通常采用相同的通信接口和相似的流量加密设置，使得产生的流量具有较少的歧视性。现有的加密流量分类方法要么需要手工制作和微调特征，要么精度较低。准确高效地进行 DApp 指纹识别仍然是一项具有挑战性的任务。在本文中，我们提出了 GraphDApp，这是一种使用图神经网络（GNN）的新型 DApp 指纹识别方法。我们提出了一种名为流量交互图（TIG）的图结构，作为加密 DApp 流的信息丰富的表示，它隐式地保留了双向客户端-服务器交互中的多个维度特征。使用 TIG，我们将 DApp 指纹识别转化为图分类问题，并设计了一个强大的基于 GNN 的分类器。我们从 1,300 个 DApp 收集了超过 169,000 个流量的真实流量数据集。实验结果表明，GraphDApp 在封闭世界和开放世界场景中的分类精度都优于其他最先进的方法。此外，GraphDApp在应用于传统移动应用分类时保持了较高的准确率。