PrismWF

PrismWF

Pretrained checkpoints for PrismWF: A Multi-Granularity Patch-Based Transformer for Robust Website Fingerprinting Attack.

PrismWF is a research prototype for multi-tab website fingerprinting. Use it only with traffic and systems for which you have authorization.

The exact train/validation/test indices are released separately in the PrismWF ARES Data Splits dataset repository.

Available Checkpoints

Each checkpoint was trained for 80 epochs with seed 2024, batch size 256, learning rate 5e-4, three multi-granularity blocks, and the complete six-channel slot-based Robust Trace Representation. Checkpoint selection uses validation MAP@K for the corresponding tab count.

Every checkpoint directory also contains the training history, exact run configuration, and a SHA-256 checksum.

Code and Environment

Source code, installation instructions, data preparation, and evaluation tools are available in the PrismWF GitHub repository. The reproduction environment uses Python 3.10.20 and PyTorch 2.4.1 with CUDA 12.1. NVIDIA A800 GPUs were used for validation, but inference is not tied to that GPU model.

git clone https://github.com/yyyyu120/PrismWF.git
cd PrismWF
conda env create -f environment.yml
conda activate prismwf
python -m pip install -e . --no-deps

Download a checkpoint with huggingface_hub:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="yuuu94/PrismWF",
    filename="checkpoints/closed_2tab/seed2024.pt",
)
print(checkpoint)

Evaluate it with the public repository:

python evaluate.py \
  --dataset closed_2tab \
  --checkpoint /path/to/seed2024.pt \
  --tabs 2 \
  --seed 2024 \
  --output results/closed_2tab/seed2024

The corresponding ARES multi-tab datasets are prepared using the scripts and split protocol documented in the source repository. Source traffic traces are not redistributed in this model repository.

Checkpoint Contents

The .pt files are PyTorch checkpoints produced by the public PrismWF training pipeline. Use the matching source revision and configuration JSON distributed beside each weight. SHA-256 files are provided so downloads can be verified before evaluation.

Intended Use

These weights support reproducibility research on authorized Tor traffic, multi-tab traffic analysis, robustness evaluation, and comparison under the released benchmark protocol. They are not intended for unauthorized monitoring or deanonymization of users.

Citation

Publication metadata will be added after the manuscript is published. Until then, please cite the manuscript title and the associated GitHub repository.