A Multi-scene Webpage Fingerprinting Method Based on Multi-head Attention and Data Enhancement

Aiming the problems of the high segmentation error of multi-tab webpage traffic and unreasonable use of mixed regions in the current Tor webpage fingerprinting methods and the poor recognition effect of webpage fingerprinting caused by lack of information on incomplete webpage traffic, a multi-scene webpage fingerprinting method based on multi-head attention and data enhancement was proposed. In the multi-tab webpage browsing scenario, sequence embedding and block division were used to transfer the original multi-tab webpage traffic sequence into a block matrix that preserves the original access order. Then, the global features of different mixed types of webpages were extracted from the block matrix based on multi-head attention for webpage recognition. In incomplete webpage browsing scenario, the original incomplete webpage traffic sequence was preprocessed to generate sequence samples. Then, incomplete sequences were interpolated through generative adversarial imputation nets to achieve data enhancement. The interpolated complete sequence was used for webpage recognition. The experiment results demonstrate that the proposed method has good performance in both webpage browsing scenarios.