Design And Implementation Of A Security Detection System For Recommendation Algorithms Based On Diversity Evaluation

As the Internet is about to enter the web 3.0 era,the way peo ple get information has been completely changed.Personalized reco mmendations have become the most popular feature for smartphone users.However,while users enjoy the convenience of accurate rec ommendations for their preferences,two problems have also arisen:information cocoon and privacy leakage.The current research on in formation cocoon faces problems such as no quantitative data suppo rt or old algorithms;and for privacy leakage in personalized recom mendation algorithms,there is no mature detection method or tool t o detect the compliance of recommendation algorithms and ensure u sers’ right to know in the issue of algorithm compliance and privac y protection as highlighted in the Administrative Regulations on Al gorithm Recommendation of Internet Information Services.Therefore,to address the above two issues,the results of the work in this pa per are mainly in the following aspects:1)A keyword summary algorithm is designed for news texts,weighted with three semantic features of TextRank,Bert and TF al gorithms,and the text similarity is calculated by the cosine vector method,which is able to explore the information cocooning degree of the recommendation list.Subsequently,a privacy usage detection method for recommendation algorithms is proposed,which utilizes an improved diversity algorithm,Android dynamic detection techniq ue and control variable method to grant different personal identifiers to the app,and detects which personal identifiers are used in pers onalized recommendations according to the changes in the diversity of the recommendation list,so as to discover the potential problems of privacy leakage among them.2)We designed and implemented an automated system for dete cting privacy issues in news apps on Android platform.In the reco mmendation list crawling module,a loop and crawling logic is desi gned to perform the crawling of a single page of news and the aut omatic crawling of the recommendation list of a category of news.In the dynamic monitoring module,the subject counts the sensitive personal identifiers and corresponding permissions,sensitive APIs,a nd uses the Frida framework to carry out the completion of dynami c monitoring and the modification function of the returned values.The results of the two modules are input into the detection algorith m model to obtain a complete detection system.3)Multiple sets of experiments are designed to prove the better performance of the content diversity evaluation algorithm proposed in this topic;the impact of recommendation category,time,user pr eference,and collected privacy on the formation of information coc oon is explored by detecting several popular news applications;the detection system implemented in this paper is able to discover the privacy problems that exist in it,and the correctness and usefulness of the detection algorithm proposed in this paper is demonstrated and gives relevant privacy recommendations.