A Research Of Malware Detection On Feature Selection Algorithm

With the development of mobile Internet,the number of mobile applications is explod-ing.As an open source operating system,Android is being customized by many vendors,and Android application markets swarm lots of malware because of the open characteristics,moreover,most of the mobile malware is found on Android platform.Therefor,the research of malware detection in Android application markets is a significant issue.Among different malware detection technologies,machine learning based malware detection is remarkable.Feature extraction and selection are two primary problems in Android malware detection with machine learning.Existing researches directly took the original extracted features or manually selected some of them as the input of machine learning algorithms,few of them involved strict feature selection algorithm for preprocessing the original features,which causes more time usage when modeling and classifying,resulting in lower accuracy and recall.To develop efficient Android malware detection frameworks,firstly we implement an automatic tool,AppExtractor,to extract original features and then we discover two phenom-ena,distribution bias and long tail effect,in original features with the help of two common feature selection algorithms.Secondly,we propose a novel feature selection algorithm Fre-quenSel.Most of the existing feature selection algorithms focus on calculating the statisti-cal importance of features to help selecting typical ones,however FrequenSel selects typical features according to compare the usage difference of features between malware and benign applications,which helps the features used in machine learning build more effective clas-sifiers and approximate accuracy,moreover the recall achieves the same level of accuracy.In the experiment containing 7972 Android applications,we get nearly 98%accuracy and recall,only taking 6.5s to analyzing and detecting each application.Combined with the demand of Android malware detection under big data circum-stance,we propose a deep learning based feature selection algorithm DBNSel,which is actually a kind of feature learning algorithms because of its deep belief network(DBN)architecture.Compared with feature selection algorithm FrequenSel,FrequenSel changes the number of features,while DBNSel changes the quality of features because the output of DBN is another representative of its input.We can make the dimensions of output much smaller than the dimensions of input through designing a proper DBN architecture.At last,we apply common machine learning algorithms to the output of DBN in order to detect Android malware.In the experiments based on the same application dataset we get 98.3%accuracy,moreover,we also achieve 99.4%accuracy and recall in an open test.Through these comprehensi've experiments,we have proved that the two feature selection algorithms proposed in this paper are suitable for the malware detection demand of Android application markets.