Research On Detection Techniques For Android Malware APP Based On Deep Learning

The proliferation of malicious apps on the Android platform not only poses a great threat to the privacy and property security of users,but also poses a major hidden danger to the national security.With the increasing number of malicious apps in Android,the analysis and research of malicious apps on Android has become a hot topic.In view of the problems such as the traditional Android malicious App detection technology cannot well identify the Android malicious App with anti-detection ability,the detection accuracy is low,and the detection method is less applied in the mobile phone or embedded and other restricted environments,this thesis studies and proposes an Android malicious App detection scheme based on deep learning.Firstly,this thesis studies and analyzes the current detection scheme of Android malicious App,understands the difficulties in the detection process of Android malicious App,the key methods and technologies used,and summarizes the shortcomings of the existing work.Then,aiming at the problem that apps with anti-detection ability cannot be well identified in the existing research,the anti-detection technology adopted by Android App is studied,and the solution is given for the former three generations of anti-detection technology,which can obtain the source code in the App and provide more comprehensive data for the study in this thesis.Then,aiming at the problem of low detection accuracy of traditional Android malicious App detection technology,an Android malicious App detection model based on convolutional neural network is constructed.In order to improve the generalization ability of the detection model,the appropriate activation function was set in the model to achieve the best performance of the classifier as far as possible,and the detection model of Android malicious App was further optimized from two aspects of experimental data set and model superparameter.At the same time,in order to solve the overfitting problem of model training,the Dropout layer was added in the neural network construction to properly abandon part of data and truly improve the detection accuracy of the model.The final generated detection model achieved about 97%accuracy.Finally,to solve the problem that the existing detection model is rarely used in mobile phones or embedded and other restricted environments,the Android malicious App detection model proposed in this thesis was implemented.The generated model was deployed to mobile phones using TensorFlow Lite technology,so that local detection of App files could be achieved.All source code developed in this thesis has been uploaded to Github for open source sharing.