ClassifyDroid: Large scale Android applications classification using semi-supervised Multinomial Naive Bayes

Rapid advances in mobile internet have enabled mobile applications to enter the era of `Big Data' (large datasets). Classification on large scale Android applications has attracted great interest from both researchers and practitioners. However, most existing approaches are supervised learning method which needs lots of labeled data. Their use in practice is often limited due to lack of labeled data, large scale Android applications, or high manual label cost. In this paper, we present a novel large scale Android applications classification tool using semi-supervised Multinomial Naive Bayes (SMNB) algorithm, called ClassifyDroid. Our proposed model exploits SMNB algorithm widely used in text document analysis. The approach is based on the analysis of characteristic application program interface (API), which can be seen as equivalents to the words and keywords in a text document. Namely, each application is characterized as a vector according to the characteristic API in it, with the associated frequencies. We evaluated ClassifyDroid on 15590 samples chosen from mobile market (MM) App Store. Our experiments show that ClassifyDroid is both accurate and practical, which has a better classification result than MNB algorithm when the dataset contains little labeled applications and lots of unlabeled applications.