AdGraph: A Machine Learning Approach to Automatic and Effective Adblocking
Umar Iqbal (The University of Iowa), Peter Snyder (Brave Software), Shitong Zhu (University of California Riverside), Benjamin Livshits (Brave Software and Imperial College London), Zhiyun Qian (University of California Riverside), Zubair Shafiq (The University of Iowa) | Machine Learning, Privacy
Filter lists are widely deployed by adblockers to block ads and other forms of undesirable content in web browsers. However, these filter lists are manually curated based on informal crowdsourced feedback, which brings with it a significant number of maintenance challenges. To address these challenges, we propose a machine learning approach for automatic and effective adblocking called AdGraph. Our approach relies on information obtained from multiple layers of the web stack (HTML, HTTP, and JavaScript) to train a machine learning classifier to block ads and trackers. Our evaluation on Alexa top-10K websites shows that AdGraph automatically and effectively blocks ads and trackers with 97.7% accuracy. Our manual analysis shows that AdGraph has better recall than filter lists, it blocks 16% more ads and trackers with 65% accuracy. We also show that AdGraph is fairly robust against adversarial obfuscation by publishers and advertisers that bypass filter lists.
