This repository contains the artifacts for the paper: "Swallow: A Transfer-Robust Website Fingerprinting Attack via Consistent Feature Learning" to appear at CCS 2025.

You can read our paper at /paper/Swallow.pdf. The published version: https://dl.acm.org/doi/10.1145/3719027.3744795.

Swallow/
├── config.yaml                  # Configuration file for all parameters
├── ProcessData.py               # Data processing implementation
├── PreTrain.py                  # Pre-training implementation
├── fineTuning.py                # Fine-tuning implementation for closed-world
├── fineTuning_open.py           # Fine-tuning implementation for open-world
├── fineTuning-AE.py             # Fine-tuning implementation for AE scenario
├── trainer.py                   # Training utilities and functions
├── utils.py                     # Utility functions and helpers
├── models/                      # Model architectures and definitions
│   ├── resnet_base_network.py   # ResNet model implementation
│   └── mlp_head.py              # MLP model implementation
├── loader/                      # Data augmentation
│   └── dataAugmentation.py      # Data augmentation file
├── SaveModel/                   # Saved model checkpoints
├── SaveData/                    # Processed datasets
│   ├── PreTrain/                # Pre-training datasets
│   └── FineTune/                # Fine-tuning datasets
├── results/                     # Experiment results and outputs
│   ├── close/                   # Closed-world evaluation results
│   └── open/                    # Open-world evaluation results
└── Traces/                      # Raw traffic traces
    ├── D1-Undefence             # Raw dataset 1 (undefended)
    ├── D2-Undefence             # Raw dataset 2 (undefended)
    └── ...                      # Additional raw trace directories

The project is compatible with the following operating systems and environments:

• Windows or Ubuntu (Linux)

• Python 3.10 is required.
• Must use the python3 command (do not use python, which may point to Python 2.x)

• CUDA-compatible GPU with appropriate drivers installed
• PyTorch-compatible CUDA version

• At least one GPU for model training and evaluation
• 100GB of available disk space to store raw traces, processed datasets, and model checkpoints

• GPU memory: At least 8GB or more (e.g., NVIDIA RTX 3080 / Tesla V100)
• RAM: 16GB or higher
• SSD storage: For faster data loading and processing

• We provide the pre-trained models and corresponding datasets for the first two scenarios in the closed-world setting, as well as the pre-trained model and data for the first scenario in the open-world setting.
• In addition, we also share some of the original traces from the scenario #1 in the closed-world setting. More details about this dataset can be found in Table 1 in the paper.
• You can download the above datasets and pre-trained models via https://zenodo.org/records/16607834.

1. Download or copy the project files to your local environment.
2. Install dependencies:

   pip install -r requirements.txt

The config.yaml file contains all necessary configuration parameters. Key paths to configure:

path:
  trace: "./Traces/"              # Path to raw traffic traces
  pretrain_data: "./SaveData/PreTrain/"  # Path to pre-training datasets
  finetune_data: "./SaveData/FineTune/"  # Path to fine-tuning datasets
  model_path: "./SaveModel/"      # Path to save models

You can directly run the trained model to obtain the corresponding results.

python3 fineTuning.py

Results are output as .txt files in the results directory, corresponding to Tables 2, 3, and 12 in the paper.

Details:

• Lines 184–192 of fineTuning.py produce ./results/close/results-Scenario#1.txt, which corresponds to Table 2 (N=5) and Table 12 (N=10,15,20) in Section 6.2 Scenario #1.
• Lines 196–202 of fineTuning.py produce ./results/close/results-Scenario#2.txt, which corresponds to Table 3 in Section 6.2 Scenario #2.

The .txt format includes:

• Format: <PretrainDataset>-<FinetuneDataset>-N=<value>
• Fields: avg (mean accuracy), std (standard deviation), max (maximum accuracy)

You can adjust the following macro parameters to reproduce specific rows in the tables:

• N = []: list of sample sizes to evaluate
• defence_type_list = []: list of defense methods to include

python3 ./results/close/txt2table.py

After completing the execution of fineTuning.py, run the above command to convert the original results.txt into a more readable tabular format.

results/close/results-Scenario#1-table.csv corresponds to Table 2 (N=5) and Table 12 (N=10,15,20) in Section 6.2 Scenario #1.

results/close/results-Scenario#2-table.csv corresponds to Table 3 in Section 6.2 Scenario #2.

python3 fineTuning_open.py

Outputs include raw precision-recall values saved as .csv files and a generated figure open_result.png in ./results/open/.
If .csv results are already present, the script will directly generate the plot.
These correspond to Section 6.3 and Figure 8 in the paper.

You also can reproduce the entire training and testing pipeline, including feature extraction from the raw Traces, pre-training, and fine-tuning.

python3 ProcessData.py

python3 PreTrain.py

python3 fineTuning-AE.py

Note: This full pipeline demonstrates feature extraction → pre-training → fine-tuning using raw trace data.
Due to dataset size, we only include traces from Datasets D1–D5 under the Undefense, WTF-PAD, and Front scenarios. These correspond to partial results in Table 2 and Table 12 (Section 6.2 Scenario #1).
As in Direct Testing of the Results, you can modify N = [] to reproduce specific rows in the tables.

The published version and the standard citation format will be updated later.

If you have any questions, please get in touch with us.

• Junyu Ai (aijunyu@bit.edu.cn)

More detailed information about the research of Meng Shen Lab can be found here (https://mengshen-office.github.io/).