MetaFETC: Adaptive Few-Shot Encrypted Traffic Classification with Meta-Learning for Diverse Encryption Techniques

Encrypted Traffic Classification (ETC) plays a pivotal role in managing network traffic and ensuring network security. Existing solutions often rely on training on large-scale traffic datasets to achieve good performance. However, due to the diversity of encryption technologies and application scenarios in the real world, collecting and labeling sufficient data for each ETC scenario often becomes impractical. To address this problem, we propose a novel adaptive few-shot ETC model named MetaFETC. MetaFETC extracts protocol-agnostic packet length sequences from encrypted traffic sessions as the input to learn the universal traffic representation and utilizes multi-task learning and meta-learning strategies to quickly adapt to new unknown ETC tasks in evolving network environments with only a small amount of new labeled data. Specifically, we utilize Temporal Convolution Network (TCN) model to capture the dependencies within the packet length sequences, which is essential for understanding the complex patterns present in encrypted traffic. Moreover, we introduce a multi-task meta-learning module to learn the optimal model initialization parameters via non-target task data of other encrypted scenarios, which can facilitate generalization and transferability across novel diverse ETC tasks. Experimental results show that the classfication accuracy of MetaFETC improves 10% in unknown few-shot encrypted scenario compared to traditional transfer method.