Skip to content

Latest commit

 

History

History

page-based

README.md

Code for training, evaluating and attacking a page-based perceptual ad-blocker.

The code in this directory used to load and run the YOLO-v3 model is inspired by tensorflow-yolo-v3.

Data Collection and Training

Our scripts for data collection can be found in data-collection.

Evaluation

A pre-trained classifier is available from https://github.com/ftramer/ad-versarial/releases The below script expects page_based_yolov3.weights to be placed under ../models. The collected web data contained in the release should be under ../data.

To evaluate the model on 20 screenshots of news websites (with outputs in a newly created directory temp), use:

python classify.py --input_dir=../data/page_based/web/test/ --output_dir=temp

Attacks

Below is a list of commands for reproducing different attacks considered in our paper. The attack samples will be saved in subdirectories under a global output directory.

Universal attacks for all webpages:

  • Evasion Attack (C4U in the paper): The publisher overlays a transparent mask on the full webpage to evade ad-blocking:
python -m attacks.web_evade_all_ads_overlay --input_dir=../data/page_based/web/
  • Test the generated mask when scrolling over a full webpage:
python -m attacks.web_test_scrolling --full_page=../data/page_based/nytimes_full.png --mask=output/overlay/mask_
  • Evasion Attack (C4U' in the paper): The publisher overlays a transparent mask on the full webpage to disable ad-blocking by overflowing it with inccorect predictions:
python -m attacks.web_overflow_all_overlay --input_dir=../data/page_based/web/mask_100.png
  • Detection Attack (C1U in the paper): The publisher adds a footer added on the bottom of a page that triggers a false ad prediction:
python -m attacks.web_footer_false_positive --input_dir=../data/page_based/web/

Attacks fine-tuned for bbc.com:

  • Evasion Attack (C4 in the paper): The publisher perturbs bottom of ad frame to evade ad-blocking:
python -m attacks.bbc_evade --input_dir=../data/page_based/bbc/
  • Evasion Attack (C3 in the paper): The ad-network perturbs ads using a universal perturbation to evade ad-blocking:
python -m attacks.bbc_evade_ad_network --input_dir=../data/page_based/bbc/
  • Detection Attack (C1 in the paper): Publisher perturbs the page header to create a false ad prediction:
python -m attacks.bbc_false_positive --input_dir=../data/page_based/bbc/

Demo of abuse attack on Facebook

Include keras-yolo3:

export PYTHONPATH=$PYTHONPATH:keras-yolo3

Convert model to Keras:

python -m keras-yolo3.convert ../models/page_based_yolov3.cfg ../models/page_based_yolov3.weights ../models/page_based_yolov3.h5

Run attack (A1 in the paper):

python -m attacks.fb_abuse