Research On Android Apps Malicious Advertising Detection Method Based On Deep Learning

The rapid development of mobile Internet has greatly driven the development of smart phones.The popularity of smart phones makes advertising placed in them have greater publicity,exposure and benefits.Many advertisers are gradually moving their campaigns from traditional formats to mobile apps in response to this trend.However some advertisers and app developers are in pursuit of more profits.They put malicious advertising in apps by various means.This behavior seriously affects the healthy development of the advertising ecosystem and raises many security and privacy issues.Therefore,how to detect malicious advertising in mobile apps has become a current research hotspot.This paper analyzes the current domestic and international methods for detecting malicious advertising in apps.Targeting malicious advertising text in apps has the characteristics of obscure rubbish,which leads to the ineffectiveness of traditional detection methods.And the previous detection method has the problem of low accuracy in detecting close induced buttons using a generic target detection algorithm.Therefore,this paper proposes a deep learning-based malicious advertising detection method for Android apps.The main findings of this paper are as follows:(1)For the problem of long time consuming advertising extraction in mobile,this paper proposes an ad＿first strategy based on the droidbot automated testing framework.In order to be able to prioritize the exploration of advertising interfaces in a shorter period of time,by collecting and analyzing the features of advertising controls and non-advertising controls in 12 types of apps.Then this paper uses the mutual information feature selection algorithm to get the set of features with high relevance and assign the corresponding weights respectively.Finally,the controls are sorted according to the weight size of the control features,and the sorted controls are operated in turn to realize the ad＿first strategy.(2)In view of the fact that malicious advertising text has the characteristics of obscure and suggestive,the traditional text detection methods are not effective.This paper proposes a hybrid neural network malicious advertising text classification model based on BERT word vector and Bi GRU-Mul Atten.The text classification model is trained by collecting advertising texts during testing and on the network platform to constitute a data set.Experiments show that the model has better detection performance than other text classification models.(3)To address the problem that previous detection methods using a generic target detection algorithm could not effectively detect the close induced buttons.This paper proposes a close induction key detection method based on improved Faster R-CNN.This paper collects and analyzes relevant advertising examples and investigates the relevant regulations of national advertising design.The current combinations of close button that may induce users to click are divided into three categories.The detection method uses from text optical character recognition and statistics of keyword word frequency to image small target detection algorithm to realize the detection of advertising close induced buttons.Experiments show that the method has a good detection effect.