An Android Malware Detection Approach Using Bayesian Inference

Android malware detection has been a popularresearch topic due to non-negligible amount of malwaretargeting the Android operating system. In particular, thenaive Bayes generative classifier is a common techniquewidely adopted in many papers. However, we found thatthe naive Bayes classifier performs badly in ContagioMalware Dump dataset, which could result from theassumption that no feature dependency exists. In this paper, we propose a lightweight method for An-droid malware detection, which improves the performanceof Bayesian classification on the Contagio Malware Dumpdataset. It performs static analysis to gather malicious fea-tures from an application, and applies principal componentanalysis to reduce the dependencies among them. Withthe hidden naive Bayes model, we can infer the identityof the application. In an evaluation with 15,573 normalapplications and 3,150 malicious samples, our work detects94.5% of the malware with a false positive rate of 1.0%.The experiment also shows that our approach is feasibleon smartphones.