Website-Fingerprinting-&-Mockingbird-Defense-Library

The following is based on the instructions provided by the developers of the WFLIB: "WFlib is a Pytorch-based open-source library for website fingerprinting attacks, intended for research purposes only.

Website fingerprinting is a type of network attack in which an adversary attempts to deduce which website a user is visiting based on encrypted traffic patterns, even without directly seeing the content of the traffic.

We provide a neat code base to evaluate 11 advanced DL-based WF attacks on multiple datasets. This library is derived from our ACM CCS 2024 paper. If you find this repo useful, please cite our paper."

The main scope of our project was to use a few of the attacker models to experiment with defended datasets, more specifically tested out with the Mockingbird Defense. We hope for future work, we extend a defensive model library on top of the pre-exisisting attacker model library.

@inproceedings{deng2024wflib,
  title={Robust and Reliable Early-Stage Website Fingerprinting Attacks via Spatial-Temporal Distribution Analysis},
  author={Deng, Xinhao and Li, Qi and Xu, Ke},
  booktitle={Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security},
  year={2024}
}

Contributions via pull requests are welcome and appreciated.

WFlib Overview

The code library includes 11 DL-based website fingerprinting attacks.

Attacks	Conference	Paper	Code
AWF	NDSS 2018	Automated Website Fingerprinting through Deep Learning	DLWF
DF	CCS 2018	Deep Fingerprinting: Undermining Website Fingerprinting Defenses with Deep Learning	df
Tik-Tok	PETS 2019	Tik-Tok: The Utility of Packet Timing in Website Fingerprinting Attacks	Tik_Tok
Var-CNN	PETS 2019	Var-CNN: A Data-Efficient Website Fingerprinting Attack Based on Deep Learning	Var-CNN
TF	CCS 2019	Triplet Fingerprinting: More Practical and Portable Website Fingerprinting with N-shot Learning	tf
BAPM	ACSAC 2021	BAPM: Block Attention Profiling Model for Multi-tab Website Fingerprinting Attacks on Tor	None
ARES	S&P 2023	Robust Multi-tab Website Fingerprinting Attacks in the Wild	Multitab-WF-Datasets
RF	Security 2023	Subverting Website Fingerprinting Defenses with Robust Traffic Representation	RF
NetCLR	CCS 2023	Realistic Website Fingerprinting By Augmenting Network Trace	Realistic-Website-Fingerprinting-By-Augmenting-Network-Traces
TMWF	CCS 2023	Transformer-based Model for Multi-tab Website Fingerprinting Attack	TMWF
Holmes	CCS 2024	Robust and Reliable Early-Stage Website Fingerprinting Attacks via Spatial-Temporal Distribution Analysis	WFlib

We implemented all attacks using the same framework (Pytorch) and a consistent coding style, enabling researchers to evaluate and compare existing attacks easily.

Usage

Install

git clone git@github.com:imakelaa/ML-WF.git
pip install --user .

Note

• Python 3.8 is required.

Datasets

mkdir datasets

• Download VPN/Non-VPN datasets (VNAT_release_1.zip (36 GB)) and unzip into ./vpn_wf_datasets/unzip_OG/

The raw dataset contains 162 pcap files across 10 traffic classes (rdp, skype-chat, youtube, netflix, rsync, scp, sftp, ssh, vimeo, voip) in both VPN and non-VPN variants. To produce the final .npz dataset used for training, the following preprocessing pipeline was applied from inside the vpn_wf_datasets/ directory:

Step 1 — Filter non-VPN captures (removes TCP noise):

# Reads from nonvpn_UNfiltered/, writes to nonvpn_captures_filtered/
python filter_nonvpn_captures.py

Applies a tshark display filter: TCP-only, strips retransmissions, fast retransmissions, spurious retransmissions, duplicate ACKs, lost segments, out-of-order packets, and previous-segment-not-captured flags.

Step 2 — Filter VPN captures (removes protocol noise):

# Reads from extra_vpn/, writes to extra_vpn_udp/
python filter_vpn_captures.py

Removes DNS and ICMP packets from VPN traffic (VPN tunneling includes mixed protocol overhead).

Step 3 — Convert filtered pcaps to NPZ (per-packet timestamps and sizes):

Using a pcap-parsing tool (e.g. dpkt or scapy), extract per-packet arrival timestamps and byte sizes from each filtered pcap. Save each capture as a .npz file with keys times (1D float array) and sizes (1D float array) into the appropriate folder:

vpn_npz_time/       vpn_npz_size/
nonvpn_npz_time/    nonvpn_npz_size/

After filtering, 3 traffic classes with sufficient packet counts were retained: rdp, skype-chat, youtube — 3 captures each for VPN and non-VPN (18 capture files total).

Step 4 — Build the combined dataset:

# Run from vpn_wf_datasets/
python build_data.py --root . --out_name new_wf_L65536 --L 65536 --num_windows 10

Pairs matching time and size npz files, concatenates them into a single 2*L = 131072-length feature vector per window, and assigns a numeric label encoding both domain (VPN=1, non-VPN=0) and class. Labels are assigned as domain_id * num_classes + class_id (e.g., non-VPN rdp=0, VPN rdp=3). Sliding-window augmentation produces multiple samples per capture file.

Step 5 — Split into train/valid/test:

python exp/dataset_process/dataset_split.py --dataset new_wf_L65536

The resulting dataset new_wf_L65536.npz has the following composition:

Traffic Class	VPN Samples	Non-VPN Samples	Total Samples	Label (non-VPN / VPN)
rdp	25	20	45	3 / 0
skype-chat	3	3	6	4 / 1
youtube	17	19	36	5 / 2
Total	45	42	87	—

• Each sample has shape (131072,) — the first 65536 values are packet timestamps, the next 65536 are packet sizes (zero-padded to length L=65536).
• The y label encodes both whether the traffic is VPN-tunneled and the application type.
• compare_size_time_npz.py can be used to verify that the times and sizes arrays for each capture have matching lengths before running build_data.py.
• view_npz.py can be used to inspect the first few values of each npz capture file.

OR

• Download TOR datasets (link) and place it in the folder ./datasets

Datasets	# of monitored websites	# of instances	Intro
CW.npz	95	105730	Closed-world dataset. Details
OW.npz	95	146446	Open-world dataset. Details
WTF-PAD.npz	95	105730	Dataset with WTF-PAD defense. Details
Front.npz	95	95000	Dataset with Front defense. Details
Walkie-Talkie.npz	100	90000	Dataset with Walkie-Talkie defense. Details

• The extracted dataset is in npz format and contains two values: X and y. X represents the cell sequence, with values being the direction (e.g., 1 or -1) multiplied by the timestamp. y corresponds to the labels. Note that the input of some datasets consists only of direction sequences.

• Divide the dataset into training, validation, and test sets.

# For single-tab datasets
python exp/dataset_process/dataset_split.py --dataset CW
# For multi-tab datasets
python exp/dataset_process/dataset_split.py --dataset Closed_2tab --use_stratify False

Training & Evaluation

We provide all experiment scripts for WF attacks in the folder ./scripts/. For example, you can reproduce the DF attack on the CW dataset by executing the following command.

bash scripts/DF.sh

Contact

If you have any questions or suggestions, feel free to contact:

• Pakhi Sinha (pasinha@ucsc.edu)
• Mahyar Vahabi (mvahabi@ucsc.edu)

Acknowledgements

We would like to thank all the authors of the referenced papers and the original developers of the attacker model library.