Enhancing Search Privacy on Tor: Advanced Deep Keyword Fingerprinting Attacks and BurstGuard Defense

[Paper Link]

Chai Won Hwang*, Hae Seung Jeon*, Ji Woo Hong, Ho Sung Kang, Nate Mathews, Goun Kim, and Se Eun Oh†

*Equally credited authors. †Corresponding author.

Note

This is the DKF attack model and BurstGuared defense proposed in Enhancing Search Privacy on Tor: Advanced Deep Keyword Fingerprinting Attacks and BurstGuard Defense work, presented in the ASIACCS'25.

1. Environment

We utilized a single NVIDIA RTX A6000 GPU (40GB VRAM) in a Ubuntu 20.04 server with 1.0 TB RAM, 7TB SATA SSDs, two NVMe SSDs, and CUDA 11.4.

2. Prerequisites and Settings

2-1. Python Dependencies

For experiments, we used the dependencies below:

tensorflow==2.6.0
keras==2.6.0
scikit-learn==1.3.0
numpy==1.22.4
pandas==2.2.2
parmap==1.7.0
tqdm==4.66.4
natsort==8.4.0

3. Dataset

For the datasets, you can use the download link below. Note that the dataset is in .txt files, in Wang and Goldberg format.

Dataset	Link	Size
Bing_2023	Link	258 classes * 1,000 instances
DuckDuckGo_2023	Link	273 classes * 1,000 instances

3. Run DKF and BurstGuard

If you want to simply run the DKF model, use model/main.py file.

python3 main.py

If you want to apply BurstGuard defense and apply the DKF/TikTok/k-FP attack model, use the auto_defense.py file.

auto_defense.py

4. Contacts

Please contact us if you have any questions about KF-tbcrawler.

• Chai Won Hwang, [email protected]
• Haeseung Jeon, [email protected]
• Se Eun Oh, [email protected]