At present, Android smart mobile terminals have been becoming more and more popular, meanwhile, due to its property of high openness, the Android platform has become the main target of the attackers. In order to effectively detect the malicious software, this paper presents an approach based on BP neural network. This approach not only considers the static features of the APK of the application, but also the running characteristics of the applications. In addition, after completing collecting features, to reduce the processing capacity and the complexity of the classification algorithm, dimension reduction technique is introduced into our approach, at the same time, comparative analysis is conducted between two dimension reduction methods. Then these features are used to train and test the classification algorithm with ten -fold cross validation. The empirical results and comparative analysis demonstrate that our approach can detect unknown Android malware accurately.