In recent years, the spread of harmful information such as pornography, phishing and violence, seriously disturbs the order of the Web, causes a series of adverse effects, and especially affects young people’s physical and mental health. Statistical learning based harmful information detection methods, the current research focus, have shown their superiority for easily adapting to newly develop harmful techniques. Feature selection is one of key factors that influence the development of Web harmful information detection system. This paper will describe a novel framework for recognizing harmful Web pages. In this framework multimodal features will be extracted and each modal feather shows the different aspect of the spam information. Based on these features, we will give a feature fusion strategy. Considering the distribution of normal and harmful websites, we investigate the use of an ensemble under-sampling classification strategy to exploit the inherent imbalance of labels in this classification problem.