AdScan: A Graph-Based Multi-modal Framework for Fine-Grained Ad Detection

Dataset: In this paper, we introduce three datasets: SITE-D, IMG-D, and TEXT-D. SITE-D consists of 14,250 domains, including 10,000 popular websites from the Tranco list and 4,250 domains from EasyList community-reported issues. We then used an automated crawler to visit all websites in SITE-D and extracted images and text to construct IMG-D (40,000 ad images, 40,000 non-ad images) and TEXT-D (10,000 ad texts, 10,000 non-ad texts). The three datasets have been provided in this folder, and they will be fully and publicly released upon the paper's acceptance. Code: All codes will be publicly released upon the paper's acceptance. The codes (in this folder) include: Rule-based detector: Folder brave_page_graph provides the modified PageGraph source code in the rendering process for the rule-based detector, which also includes code for browser calling image-based detector and text-based detector. Folder Rule_Based_Detector provides an implementation of the rule-based detector in the detection process. Image-based detector: Folder Image_Based_ Detector provides the training and evaluating codes. Text-based detector: File Text_Based_Detector.py provides the training and evaluating codes. LabelSpread algorithm: File LabelSpread.py provides implementation of LabelSpread algorithm. Why not an online demo? The detectors are designed to integrate with the Brave browser, but uploading the entire Brave source code (over 60GB) is impractical. Instead, we are preparing a patch that allows readers to download the Brave source code, apply our modifications, and compile the browser. Structure of this folder: Supplementary Materials ├─README.md ├─Supplementary Materials.pdf: Additional implementation details for paper, including the design of the rule-based detector, et al. ├─code: Implementation of individual detectors and LabelSpread. │ ├─LabelSpread.py: Implementation of LabelSpread algorithm. │ ├─Text_Based_Detector.py: Implementation of text-based detector in the detection process. │ ├─Image_Based_Detector: Implementation of image-based detector in the detection process. │ ├─Rule_Based_Detector: Implementation of rule-based detector in the detection process. │ └─brave_page_graph: Modified implementation of PageGraph in rendering process for communicating with detection process. └─data: SITE-D, TEXT-D, and a part of IMG-D. ├─SITE-D ├─TEXT-D └─IMG-D(part)