【摘要】 攻击者利用恶意社交机器人窃取用户隐私、传播虚假消息、影响社会舆论,严重威胁了个人信息安全、社会公共安全,乃至国家安全。攻击者还在不断引入新技术实施反检测。恶意社交机器人检测成为在线社交网络安全研究的一个重点和难点。首先回顾了当前社交机器人的开发与应用现状,接着对恶意社交机器人检测问题进行了形式化定义,并分析了检测恶意社交机器人所面临的主要挑战。针对检测特征的选取问题,厘清了从静态用户特征、动态传播特征,以及关系演化特征的研究发展思路。针对检测方法问题,从基于特征、机器学习、图论以及众包4个类别总结了已有检测方案的研究思路,并剖析了几类方法在检测准确率、计算代价等方面的局限性。最后,提出了一种基于并行优化机器学习方法的恶意社交机器人检测框架。更多还原

【Abstract】 The attackers use social bots to steal people’s privacy,propagate fraud messages and influent public opinions,which has brought a great threat for personal privacy security,social public security and even the security of the nation.The attackers are also introducing new techniques to carry out anti-detection.The detection of malicious social bots has become one of the most important problems in the research of online social network security and it is also a difficult problem.Firstly,development and application of social bots was reviewed and then a formulation description for the problem of detecting malicious social bots was made.Besides,main challenges in the detection of malicious social bots were analyzed.As for how to choose features for the detection,the development of choosing features that from static user features to dynamic propagation features and to relationship and evolution features were classified.As for choosing which method,approaches from the previous research based on features,machine learning,graph and crowd sourcing were summarized.Also,the limitation of these methods in detection accuracy,computation cost and so on was dissected.At last,a framework based on parallelizing machine learning methods to detect malicious social bots was proposed.更多还原