YOPO

YOPO is an attack framework designed to generate a universal adversarial perturbation that deceives machine learning-based ATS blockers in a cost-effective manner. YOPO acheives high attack success rates agaisnt four state-of-the-art ATS blockers with a single perturbation. For more details, please refer to our paper, "You Only Perturb Once: Bypassing (Robust) Ad-Blockers Using Universal Adversarial Perturbations", which appeared in ACSAC 2024.

Requirements

YOPO requires a Linux machine with an NVIDIA graphic card. Due to its use of multiple processes for parallel HTML file manipulation, we recommend a machine with at least 32 CPU cores, 256 GB of system memory, and 300 GB of disk space. Our tests were conducted on a machine running Ubuntu 22.04 (64-bit).

Installation

To run our scripts, the following software dependencies must be installed:

1. Docker

   Install Docker by following the instructions here.

2. Run Docker without root privileges

   Our scripts are designed to avoid requiring root privileges. To manage Docker as a non-root user, follow the instructions here.

3. CUDA Toolkit 12.4

   YOPO is tested with CUDA Toolkit 12.4. Please install CUDA Toolkit 12.4 from this link. If using a different CUDA version, update the base image in the Dockerfile accordingly. Available base images can be found here.

4. NVIDIA Container Toolkit

   To enable GPU access within Docker containers, install the NVIDIA Container Toolkit by following the guide here.

Usage

Step 1: Launch the Docker container

Build the Docker image using the provided Dockerfile and start a Docker container.

$ git clone --recurse-submodules https://github.com/WSP-LAB/YOPO.git
$ cd scripts
$ ./build.sh && ./launch_container.sh

Step 2: Install the mitmproxy CA certificate

Assuming you are inside the Docker container, install the CA certificate for mitmproxy, which we use to monitor requests to target websites and redirect the response to our local server.

$ cd scripts
$ ./apply_mitmproxy_certificates.sh

Step 3: Save HTML files

Before crawling webpages, save the HTML files from Tranco's Top-10K websites.

$ cd scripts/crawler
$ ./save_html.sh

Step 4: Crawl websites

Crawl the websites using the target ATS blockers by executing the run_crawler.sh script, passing the name of the target ATS blocker. Each ATS blocker crawls the saved HTML files and extracts features.

• TARGET: adgraph, webgraph, adflush, pagegraph

$ cd scripts/crawler
$ ./run_crawler.sh [TARGET]

Step 5: Train ATS blockers

Train the target ATS blockers with the extracted features by running the train_adblocker.sh script.

$ cd scripts
$ ./train_adblocker.sh [TARGET]

Step 6: Train surrogate models

Train a surrogate model for each target ATS blocker by executing the train_surrogate.sh script.

$ cd scripts
$ ./train_surrogate.sh [TARGET]

Step 7: Perform the attack

To run the attack pipeline, execute the run_attack.sh script with the following arguments:

• EPSILON: 5, 10, 20, 40
• COST_MODEL_TYPE: DC, HSC, HCC, HJC
• TARGET: adgraph, webgraph, adflush, pagegraph

$ cd scripts
$ ./run_attack.sh [EPSILON] [COST_MODEL_TYPE] [TARGET]

The attack success rates and costs will be saved in the /yopo-artifact/result directory.

Authors

This research project has been conducted by WSP Lab at KAIST and SAIL Lab at Oregon State University.

• Dongwon Shin
• Suyoung Lee
• Sanghyun Hong
• Sooel Son

Citation

To cite our paper:

@INPROCEEDINGS{shin:acsac:2024,
  author = {Dongwon Shin and Suyoung Lee and Sanghyun Hong and Sooel Son},
  title = {You Only Perturb Once: Bypassing (Robust) Ad-Blockers Using Universal Adversarial Perturbations},
  booktitle = {Proceedings of the Annual Computer Security Applications Conference},
  pages = {190--206},
  year = 2024
}