Identifying Gambling Websites with Co-training

Abstract

Gambling websites do great harm to society and many even cause serious network crime. To identify the gambling websites, many machine learning based methods are proposed by analysing the URL, the text, and the images of the websites. Nevertheless, most of them ignore one important information, i.e., the text within the website images. The text on the images of gambling websites has keywords that clearly point to such websites. Motivated by this, in this paper, we propose an co-training based gambling website identification method by combining the visual and semantic features of the website screenshots. First, we extract text information from webpage screenshots through the optical character recognition (OCR) technique. Then we train an image classifier based on a convolutional neural network (CNN) and a text classifier based on TextRNN respectively from image view and text view. Second, the two classifiers are retrained on unlabeled data with the co-training algorithm. Third, we conduct experiments on the webpage screenshot dataset we collected. The experimental results indicate that OCR text has strong semantic feature and the proposed method can effectively improve the performance in identifying gambling websites.