Research On Android Malicious Software Detection Based On Deep Learning

The mobile Internet has developed rapidly in recent years,and the Android smart phone operating system has gradually become the most popular mobile terminal operating system.The software of the Android operating system can be used freely and it is popular with mobile users,so it has become the key target for malicious viruses.Signature detection is the mainly detection method for Android malicious appl-ications,this detection method has low detection ability for unknown malicious softwares and has a high false alarm rate.Aiming at the problem,thesis implements a malicious application detection method based on deep learning from the view of static analysis and machine learning.Basing on the detection method,an unknown Android malicious application detection model is trained.The main contributions are summarized as follows:Firstly,the normal applications and malicious applications are extracted as training set.The source code of the application is obtained through static analysis,and the API(Application Programming Interface)is extracted from the source code as training feature.Secondly,the API features are reduced through the depth AutoEncoder network,and then the Logistic Regression Binary Detection Model is trained.This model can detect whether an unknown Android application is malicious.Finally,in order to achieve more fine grained malicious application detection,thesis uses four malicious familys to train the Softmax multiclassification model which is able to detect the malicious family category.The detection process as follows:first,binary classification tests are performed on the application through the logistic regression model;Se-cond,if the application is detected as the malicious category,the Softmax multi-classification model is used to detect the malicious family catego-ry.Through separating test sets to test the detection effect of the model,the test result proves that Android malicious application detection model based on deep learning is well effective.The recall rate of the model is 95.2%.