Cerberus:基于深度学习的跨网站社交机器人检测系统  

Cerberus:cross-site social bot detection system based on deep learning

作　　者：汤家伟 刘育杉 高敏 宫庆媛 王新[1] 陈阳[1] TANG Jiawei;LIU Yushan;GAO Min;GONG Qingyuan;WANG Xin;CHEN Yang(School of Computer Science,Fudan University,Shanghai 200438,China;Research Institute of Intelligent Complex Systems,Fudan University,Shanghai 200438,China)

机构地区：[1]复旦大学计算机科学技术学院,上海200438 [2]复旦大学智能复杂体系基础理论与关键技术实验室,上海200438

出　　处：《智能科学与技术学报》2024年第4期480-494,共15页Chinese Journal of Intelligent Science and Technology

基　　金：国家重点研发计划(No.2022YFB3102901);国家自然科学基金项目(No.62072115,No.62102094);上海市科技创新行动计划(No.22510713600)。

摘　　要：社交网站吸引了数十亿用户,影响着人们的生活方式。社交网站作为开放平台,注册加入的门槛较低,社交机器人能够轻易地注册,并进行舆论导向控制、不实信息传播等有害活动,以谋取利益。单一社交网站的机器人检测系统往往需要依赖用户的历史行为数据进行分析。因此,社交机器人在被识别出之前往往已经成功实施了恶意攻击。为尽早地识别出社交机器人,提出了跨网站社交机器人检测系统Cerberus。Cerberus可以解决用户早期在单个社交网站上数据不充足导致的用户识别“冷启动”的问题。Cerberus使用用户在Medium网站上的个人信息和历史活动信息,对用户链接在Twitter上的账号是否为社交机器人账号进行预测。结果表明,该系统的AUC值可达0.7552,具有良好的识别准确性。Social networking sites have attracted billions of users and influence people’s lifestyles.However,as open plat‐form with low requirements for registration and joining,it is inevitable that social bots are able to easily register and do harmful things such as controlling public opinions and spreading inaccurate information for profit.Nevertheless,single-site social bot detection systems often rely on historical behavioral data to identify bots,and the detection occurred after the social bots have implemented their attacks.To identify social bots as early as possible,this paper proposed Cerberus,a cross-site social bot detection system.Cerberus can solve the cold-start problem of user identification caused by insuffi‐cient user data on a single platform at an early stage.Cerberus used personal information and historical activity on the Me‐dium website of users to make prediction about whether a user’s account on Twitter was a social bot.The results from our experiments show that the AUC score of Cerberus can reach 0.7522,which has good recognition accuracy.

关 键 词：在线社交网络 社交机器人检测 跨网站链接 深度学习 冷启动用户 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]