WSNet: A Wrapper-Based Stacking Network for Multi-scenes Classification of DApps.

Decentralized applications (DApps) are growing rapidly with the prevalence of blockchain, but security and performance issues plague network managers and developers. Encrypted network traffic classification (ETC) plays a fundamental role in application management, security detection, and QoS improvement and requires different granularity for different scenarios. Existing work focuses on a single scenario, and objects of them are traditional centralized applications (Apps). Since DApps use similar encrypted traffic settings and the same communication interface, the traffic is more complex than Apps. Under the premise of manual-design features, sophisticated architecture, or lots of training data, existing methods have good results, otherwise suffering from low accuracy. In this paper, we propose Wrapper-based Stacking Network (WSNet). According to traffic characteristics of different scenarios, WSNet adaptively selects optimal features for different algorithms to filter out irrelevant and redundant features without human intervention, thereby improving classification efficiency. Combining with stacking technology to integrate advantages of primary learners, hence it has good performance in complex traffic scenarios. Our comprehensive experiments on two real-world datasets show that WSNet adapts to and outperforms the state-of-the-art methods.