Research On Android Malicious Code Injection Detection Technology Based On Semantic Features

With the rapid development and popularity of Android smart phones,the number of devices and software based on Android system is also exploding.Compared with Apple’s IOS system,the open source Android system is more vulnerable to attack,and malicious code can be injected and attacked the mobile terminal easily.At present,the research work in the detection of Android malicious applications mainly focuses on how to better extract the features of Android applications.Among these features,the most reliable method is to use semantic features to judge malicious behaviors.In order to extract the semantic features of applications more accurately,this paper USES the static analysis method to extract the semantic features of Android applications,and selects the key features based on the attention mechanism,and proposes an improved integrated learning classifier to detect code injection to attack malicious applications.The specific work of this paper is as follows:First,the abstract syntax tree is used as the code representation to traverse the node to get a piece of text containing the code details.Then,the n-gram model in natural language processing is used to make the feature sub-sequence originally composed of a single word become composed of several words.Secondly,comparing the code injection detection with the text classification in the field of natural language processing,the attention-based Semantic Feature(ASF)model is designed,which abstracts the Feature of the source text through the Attention mechanism in the neural network.Thirdly,considering the comprehensive performance,a random forest classifier was selected for Android malicious code detection,and a random forest algorithm Based on Accuracy and Correlation(ACRF)was proposed to improve the classification Accuracy of the algorithm.Finally,the proposed detection model is analyzed experimentally.Compared with the manual feature selection and the partial neural classification algorithm with long training time,the proposed semantic feature extraction based on Attention mechanism and the improved random forest detection model have significant advantages in the final classification accuracy and convergence rate.