With the rapid development of the Web, Internet gambling has become a global problem, which causes nontrivial social impacts. Despite of its prosperity, in the major countries such as United States, Russia, and mainland China, Internet gambling is explicitly prohibited, and in the most remaining countries, Internet gambling is under strict regulations. However, there are so many websites that it is rather difficult to regulate Internet gambling and rather challenging to identify them. It may introduce many false positives or false negatives, if we simply grep contents of websites with keywords. In this paper, we find that the behavior of HTTP POST is a strong indicator to detect gambling sites. Based on the finding, we propose a novel approach that detects gambling sites with mined behavior models of such sites. Furthermore, we introduce graph analysis to improve our approach. Our evaluation shows that our approach achieves high precision and recall, when it detects online gambling sites from a large number of websites.