Research On Deep Learning Algorithm Of Android Malware Static Detection Based On Behavioral Pattern

The openness of the Android operating system makes it the most popular smartphone operating system in the world.However,this openness has also attracted many hackers to develop and install malware.These malware have caused serious economic losses and privacy leaks to users for the purpose of tariff consumption,privacy theft,malicious deduction,and remote control.Therefore,accurate and fast detection of malware has become a huge challenge for the Android market,where a large number of applications are online and updated every day.At present,there are many algorithms to manually extract features from Android applications to achieve the purpose of detecting Android malware.However,malware developers are always able to find an endless stream of code obfuscation methods that make the accuracy of detection algorithms much lower while maintaining application maliciousness.In view of the above problems,this thesis studies the static detection algorithm of Android malware from the perspective of behavioral mode and deep learning technology.This thesis presents a design a behavioral pattern based static detection deep learning the algorithm for Android malware.The algorithm includes feature extraction for Android decompiled code,digital vector coding for feature features,and multi-layer bidirectional LSTM neural network based on dense attention layer.Focusing on these three parts,this thesis first studies how to extract the behavior pattern features that can reflect the malicious behavior of Android applications and the attribute characteristics of Android application static attributes from Android decompiled code.Then,this thesis proposes an encode scheme base on the similarity of behavior pattern.The scheme uses Word2vec coding method for digital vector coding.According to the mutual independence of attribute features and the finiteness of the dimension,the One-Hot coding method is used for digital vector coding.After that,this thesis designs a dense attention mechanism based on the attention layer.At the same time,the multi-layer bidirectional LSTM neural network is used to explore the logical relationship between the behavior pattern features to make the detection more accurate.In this thesis,the effects of different configurations of various parts of the proposed algorithm on the test results are analyzed by ablation experiments.The algorithm proposed in this thesis experiments on 23,200 Android application sample sets,achieving an accuracy of 97.7%,and the average detection time of each Android application is only 1.36 seconds.Finally,the experimental comparison of the methods this year shows that the proposed algorithm is not only superior to a similar algorithm in accuracy,but also has an improvement in detection speed.