<p>The development of distributed networks has increased the number of communication nodes, heightening the risk of private information leakage. Tor networks mitigate this risk through multi-hop routing and multi-layered encryption, providing anonymous communication. However, attackers eavesdrop on the traffic and analyze patterns like packet size and direction to identify the website the user visited. This is known as website fingerprinting (WF) attacks, and related research has contributed to defenses, thereby protecting privacy in distributed networks. In recent years, learning-based approaches have significantly enhanced their effectiveness by automatically extracting complex features. Nevertheless, collecting large-scale data for training remains costly for attackers. Existing techniques attempt to alleviate this issue by few-shot learning (FSL), which utilizes pretraining and fine-tuning with limited new data. However, previous works primarily extract features with a general view of raw data instances, while hierarchical features, i.e., partial and general features within data classes, are overlooked. To address the limitation, we propose a novel few-shot WF (FSWF) method that consists of two stages: pretraining and few-shot fine-tuning. Our approach specifically focuses on learning equivariant and invariant features. During pretraining, a highly generalizable feature extractor is learned by introducing additional equivariant and invariant features learning tasks. This is followed by a few-shot fine-tuning using newly collected data. Our method achieves promising experimental results on open datasets.</p>