Research On Interactive Human-machine Differentiation Method Based On Content Correlation

CAPTCHA techniques provide means to differentiate humans from machines automatically.CAPTCHA has a wide range of applications in network protection and information security,such as anomaly recognition and defense against computer-based automatic attacks and resource abuse.The design of CAPTCHA,based on the open problems in the field of Artificial Intelligence(AI),makes effective use of the gap between humans and machines to solve such problems and to avoid the abuse of network services by malicious machine programs.Therefore,the identification and anti-identification of CAPTCHA is fundamentally the confrontation between CAPTCHA generation technology and artificial intelligence technology.In recent years,with the evolution of artificial intelligence technology and the reinforcement of network attack methods,CAPTCHA is facing increasing severe challenges.In view of the characteristics and problems of CAPTCHAs,this thesis mainly studies the method and implementation of human-machine discrimination based on content correlation.The major work and achievements include the followings:(1)CAPTCHA construction method based on image content correlation.Since inferencing the correlation between images is an AI-hard problem,this thesis designs a CAPTCHA method based on the correlation of image content,which imitates the divergent thinking of human beings and uses the potential associative significance of image content to design CAPTCHA.The proposed method first uses common words and mostly searched keywords to generate dynamic tag lexicon,then calculates semantic relevance between tags and constructs semantic relationship graph.A CAPTCHA generation algorithm is programed based on the random walk model on the graph,and the dynamic image resource library is created by real-time retrieval of tags on the Internet in the end,thereby avoiding the generation of regular CAPTCHA generation patterns,and for constructing highly random CAPTCHA.(2)Interaction measurement method based on user behavior process.In the procedure of CAPTCHA verification,users generate plenty of behavior characteristics data,including the order of answer clicks,the speed of mouse wandering,the trajectory of mouse movement,the time interval of clicking,and so on.This paper collects above data and carries out cluster analysis,establishes an identity authentication module based on user's behavior characteristics,scoring the user's behavior characteristics,and taking it as a hint of human-machine discrimination.(3)The construction method of CAPTCHA based on decoy effect.By adapting the concept of decoy effect,this thesis proposes a novel image-based CAPTCHA,which requires users to select several pictures related to the topic from a group of choices.CAPTCHA provides a set of pictures related to the topic,including Targets and Competitors.On the basic of the original pictures,adding "decoy" results in the consequence that human users have reinforced tend to choose the target option and achieves the purpose of distinguishing human and machine.The usability and robustness of the proposed methods have been evaluated by both numerical analysis and empirical evidence.The results show that humans can solve the CAPTCHA tests conveniently and effectively with a high pass rate,while bot-programs succeed with slim chance.The methods proposed by this paper have good usability and high security,and put forward new AI hard problems,which may also promote the development of pattern recognition.Therefore,this research is supposed to have technical significance and broad application prospects.