In this repository, we provide the code to reproduce the results in the "Understanding-and-Explaining-Web-Fingerprinting-with-a-Protocol-Centric-Approach" paper.
The repository includes reference Machine Learning models for evaluation in the models folder and tools for generating HTTPS datasets in the crawlers folder.
The code was tested using Linux Mint 21.2 Victoria and Python 3.10.
This repository is organized as follows:
models/
|- src/ # Models and evaluation methods
|- tests/ # Unit tests for the ML models
crawlers/
|- src/ # Traffic crawling and parsing
|- tests/ # Unit tests for the crawling logic
experiments/
|- domains_experiments/ # Domain experiments
|- crawler/ # Domain crawling logic
|- scripts/ # Domain fingerprinting evaluation
|- page_wiki_experiments/ # Wikipedia experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation
|- page_9gag_experiments/ # 9GAG experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation
|- page_imdb_experiments/ # IMDB experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation The evaluation models are located in the models, and they are organized in a standalone library.
The library can be installed using
cd models/
pip install -e .
pip install -e .[testing] # for the development setupXGBoost
# XGBoost usage and evaluation example
from sklearn.datasets import load_iris
# tls_fingerprinting absolute
from tls_fingerprinting.models.base.static.xgb import XGBoostClassifier as model
from tls_fingerprinting.utils.evaluation import evaluate_classifier
test_plugin = model()
X, y = load_iris(return_X_y=True, as_frame=True)
scores = evaluate_classifier(test_plugin, X, y)
print(scores["str"])
# Example Output
# {'aucroc_ovo_macro': '0.9832 +/- 0.004', 'aucroc_ovr_micro': '0.9841 +/- 0.008', 'aucroc_ovr_weighted': '0.9832 +/- 0.004', 'aucprc_weighted': '0.9766 +/- 0.005', 'aucprc_macro': '0.9766 +/- 0.005', 'aucprc_micro': '0.9766 +/- 0.005', 'accuracy': '0.9333 +/- 0.021', 'f1_score_micro': '0.9333 +/- 0.021', 'f1_score_macro': '0.933 +/- 0.021', 'f1_score_weighted': '0.9331 +/- 0.022', 'kappa': '0.9 +/- 0.032', 'kappa_quadratic': '0.9501 +/- 0.016', 'precision_micro': '0.9333 +/- 0.021', 'precision_macro': '0.933 +/- 0.021', 'precision_weighted': '0.9333 +/- 0.021', 'recall_micro': '0.9333 +/- 0.021', 'recall_macro': '0.9334 +/- 0.021', 'recall_weighted': '0.9333 +/- 0.021', 'mcc': '0.9002 +/- 0.032'}Neural Nets
# MLP usage and evaluation example
from sklearn.datasets import load_iris
import numpy as np
from tls_fingerprinting.models.base.nn.mlp import MLP as model
from tls_fingerprinting.utils.evaluation import evaluate_classifier
X, y = load_iris(return_X_y=True, as_frame=True)
test_plugin = model(
task_type="classification",
n_units_in=X.shape[1],
n_units_out=len(np.unique(y)),
)
scores = evaluate_classifier(test_plugin, X, y)
print(scores["str"])
# Example Output
# {'aucroc_ovo_macro': '0.9791 +/- 0.012', 'aucroc_ovr_micro': '0.9672 +/- 0.024', 'aucroc_ovr_weighted': '0.9787 +/- 0.013', 'aucprc_weighted': '0.9496 +/- 0.034', 'aucprc_macro': '0.9496 +/- 0.034', 'aucprc_micro': '0.9496 +/- 0.034', 'accuracy': '0.8667 +/- 0.087', 'f1_score_micro': '0.8667 +/- 0.087', 'f1_score_macro': '0.8559 +/- 0.104', 'f1_score_weighted': '0.8555 +/- 0.104', 'kappa': '0.8008 +/- 0.13', 'kappa_quadratic': '0.9081 +/- 0.052', 'precision_micro': '0.8667 +/- 0.087', 'precision_macro': '0.9025 +/- 0.039', 'precision_weighted': '0.9038 +/- 0.036', 'recall_micro': '0.8667 +/- 0.087', 'recall_macro': '0.8685 +/- 0.085', 'recall_weighted': '0.8667 +/- 0.087', 'mcc': '0.8235 +/- 0.098'}pytest -vvsxThe crawlers folder contains scripts for generating and parsing PCAP files from lists of URLS.
cd crawlers
pip install -e .
pip install -e .[testing]The experiments use a custom Selenium Docker image, with additional scripts and features. Tu build the images, run
cd crawlers/docker
docker build --tag selenium-chrome -f Dockerfile_chrome .
docker build --tag selenium-firefox -f Dockerfile_firefox .See dataset crawlers.
If the library and docker builds worked, the unit tests should pass
cd crawlers
pytest -vvsxThe experiments are available in the experiments folder. Each experiment includes the crawling scripts and the fingerprinting evaluation code. The 9GAG and IMDB are not included in the repository due size.
experiments/
|- domains_experiments/ # Domain experiments
|- crawler/ # Domain crawling logic
|- scripts/ # Domain fingerprinting evaluation
|- page_wiki_experiments/ # Wikipedia experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation
|- page_9gag_experiments/ # 9GAG experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation
|- page_imdb_experiments/ # IMDB experiments
|- crawler/ # Wikipedia crawling logic
|- scripts/ # Wikipedia fingerprinting evaluation If you use this code, please cite the associated paper:
@inproceedings{cebere2024understanding,
title={Understanding Web Fingerprinting with a Protocol-Centric Approach},
author={Cebere, Bogdan and Rossow, Christian},
booktitle={Proceedings of the 27th International Symposium on Research in Attacks, Intrusions and Defenses},
year={2024}
}