A Website Fingerprinting Attack Based on Parameter Prediction Model for Conceptual Drift

Website content updates and changes in the network environment can lead to concept drift, significantly impacting the performance of existing website fingerprinting methods. Recent research has focused on transfer learning-based approaches for concept drift adaptation to address this issue with limited samples. These methods involve pre-training a robust feature extractor and fine-tuning the classifier with a few samples. However, the effectiveness of these classifiers is highly dependent on the distributional gap between pre-training and fine-tuning datasets. As the divergence increases, more fine-tuning data is required to maintain classification performance. Consequently, existing methods struggle to achieve effective website identification tasks under severe concept drift with limited tuning data. To overcome these limitations, we propose a model parameter prediction approach, which adapts to new domains from data and model perspectives to enhance website identification performance. The proposed method predicts future model parameters to mitigate the effects of distributional changes, effectively adapting to evolving network environments. Experimental results demonstrate that our approach achieves significant improvements in accuracy, particularly under scenarios with limited tuning data and substantial concept drift.