ResAware Improves Cross-Environment Website Fingerprinting Robustness

Website Fingerprinting (WF) attacks, while highly accurate in controlled lab settings, often experience significant performance degradation in real-world environments. This degradation is attributed to factors like spatio-temporal drift, browser variations, and proxy obfuscation, primarily because these attacks rely solely on noisy and environmentally sensitive low-level traffic features. To overcome this limitation, researchers propose ResAware, a cross-environment resource-aware distillation framework. This framework operates under an asymmetric setting where training is rich in information, but inference is resource-poor. Specifically, a teacher model is trained using privileged resource-level features, which provide a deeper understanding of website characteristics. The knowledge gained by this teacher model is then distilled into a student model through heterogeneous knowledge distillation. At deployment, the student model performs inference using only encrypted traffic, incurring no additional online observation costs. Evaluations on a large, five-month dataset collected from six global vantage points demonstrate that ResAware substantially improves the cross-environment robustness of various WF baselines, showing significant F1-score and TPR@1%FPR gains even under substantial temporal drift.