Source code and dataset for our paper "Beyond Single Tabs: A Transformative Few-Shot Approach to Multi-Tab Website Fingerprinting Attacks" (WWW’25) and Its Journal Extension.

We have released the dataset associated with the novel defense method proposed in our journal extension. You can download the dataset via Google Drive link. In addition, we have made the core logic of the TGTO defense publicly available. The implementation of the TGTO obfuscation mechanism can be found in the FMWF-src directory, specifically in the notebook tgto_core_obfuscation.ipynb.

Basic Environment:

• NVIDIA GPU: RTX 3080 Ti (12GB)
• Ubuntu 22.04
• Python Version: 3.12
• CUDA Version: 12.1
• PyTorch Version: 2.3.0

In addition, you can use pip to install basic packages such as numpy, matplotlib, pandas, pyautogui, and scapy.

Source Code's Description:

The source code mainly consists of three parts:

• Traffic Capture and Processing

We used the codes 2tab-capture.ipynb, 3tab-capture.ipynb, 4tab-capture.ipynb, and 5tab-capture.ipynb from the FMWF-src folder to capture multi-label traffic data in .pcap format from Google Chrome and Microsoft Edge browsers. The captured traffic data was then processed into .csv format traffic sequences using the codes 2tab-dataprocess.ipynb, 3tab-dataprocess.ipynb, 4tab-dataprocess.ipynb, and 5tab-dataprocess.ipynb.

Our Tor traffic data was collected using the same methodology as in the paper “Robust Multi-tab Website Fingerprinting Attacks in the Wild” (published in IEEE S&P 2023).

• Pre-training with Augmented Traffic Sequences

The PretrainingModel.py file in the folder FMWF-src is used to obtain the pre-training model using manually synthesized multi-tab datasets.

• Few-shot Fine-tuning and Testing

For the Few-shot Fine-tuning and Testing section, we use the 5tab scenario as an example. The tor-5tab-Accuracy.py file in the FMWF-src folder is used to test the real-world collected 5tab traffic. This code includes both the Few-shot Fine-tuning for Real-world Adaptation and Testing on Dynamic Traffic stages for the 5tab scenario. You can reproduce our experiments using the fine-tuning and testing datasets from the tor dataset. Please note that the paths in the code may not match your environment, so be sure to replace them accordingly.

The openworld.py in the FMWF-src folder is the code used in the open-world scenario, which corresponds to the open-world experimental part of the paper.

Dataset Description:

We collected Tor single-tab and multi-tab browsing datasets under real browsing scenarios and our dataset is about 11G in size. You can download the dataset via Google Cloud Drive link. Our datasets are in .csv format. you can load them using the following code:

import pandas as pd
df = pd.read_csv('../tor5tab_dataset.csv', header=None)

Each row in the dataset represents a traffic instance. For the single-label dataset, the first item in each row is the index, the second is the repetition count, and the subsequent items are the traffic data. In the multi-label dataset, the first 100 items of each row are the multi-hot encoding of the traffic, followed by the index of the corresponding monitored website, and the remaining items are the traffic data.