Joint QUIC-TCP Website Fingerprinting

Dataset Overview

This dataset was gathered by downloading the index pages of around 16,000 websites from the Alexa Top 1M, Cisco Umbrella, and Majestic Million lists, using the Chromium web-browser, and over encrypted wireguard tunnels with gateways in New York, U.S.A.; Frankfurt, Germany; and Bengaluru, India. The traces, which were collected at the client, contain encrypted wireguard packets.

Task Description

The task is to identify the visited website given traces which may have been collected using HTTP/QUIC or using HTTP/TCP.

Links and Facts

• Dataset Link: Due to size, available upon request.
• Dataset Size (Uncompressed): 69 GB
• Disallowed Features: None
• Number of Classes: 101
• pcapML Metadata Comment Format: sampleID,final URL;original URL;tcp or quic;VPN location
• Protocols: Ethernet, IP, UDP, Wireguard
• Metric to Optimize: F_1,r-score (see notes below)

Special Dataset Notes

The dataset contains more than 100 samples per QUIC and TCP per monitored class, around monitored 100 classes, and 3 samples per QUIC and TCP per unmonitored website. The original evaluations were run on a subsampled dataset of exactly 100 samples per QUIC and TCP monitored class and 100 monitored classes, and exactly 3 samples per protocol per unmonitored website. See the code from the original paper for dataset cleaning and filtering.

The F_1,r-score is the F₁-score calculated using the r-precision score of Wang et al. (2020). It is calculated as

The F_1,r-scores below were calculated with an r-value of 20.

Citation(s)

@InProceedings{smith2021website,
    author = {Jean-Pierre Smith and Prateek Mittal and Adrian Perrig},
    title = {Website fingerprinting in the age of {QUIC}},
    booktitle = {Proceedings on Privacy Enhancing Technologies (PoPETs)},
    year = 2021,
    month = jul,
    doi = {10.2478/popets-2021-0017},
    keywords = {privacy},
}

Leaderboard