<p>The growing demand for internet privacy protection has led to the increasing popularity of anonymous communication systems such as Tor, which, however, continue to face persistent threats from website fingerprinting (WFP) attacks. Although deep learning-based WFP methods have shown promising results, existing supervised approaches still rely heavily on large labeled datasets and remain vulnerable to defense-induced noise, while also exhibiting limited generalization in few-shot scenarios due to overfitting. To address these challenges, we propose enhanced self-supervised learning for website fingerprinting attacks. Specifically, we introduce a novel framework, named Camouflage, that jointly tackles label scarcity and defense robustness through self-supervised learning. In this approach, we adopt a BYOL-based pre-training strategy to learn robust traffic representations from large-scale unlabeled data, effectively decoupling representation learning from downstream classification tasks. In addition, we incorporate an Efficient Channel Attention module into the backbone network to mitigate defense-induced noise and enhance feature discrimination, enabling high recall while maintaining low false positive rates. The proposed method is evaluated on a real-world dataset containing over 100,000 website traffic samples, covering three typical defense scenarios: No defense, WTF-PAD, and Front. Experimental results demonstrate that in a closed-world 5-shot few-shot setting, Camouflage achieves an accuracy of 91.2%. In open-world rejection tasks, it attains a high recall rate of 86.9% while maintaining an extremely low false positive rate.</p>