社交媒体已成为人们日常生活和社会交往中必不可少的一部分，社交媒体作为一种新型互联网应用形式，具有广泛的用户群体和巨大的商业价值。但是，社交媒体中存在着大量的社交机器人，严重危害了用户信息安全和社交媒体健康发展。为了净化网络空间环境，减少社交机器人的威胁，亟需使用有效的检测方法发现社交媒体中存在的社交机器人，这对社交媒体安全具有重要的意义。

随着社交机器人的迭代进化，社交机器人“拟人化”程度越来越高，导致其检测效率越来越低。现有社交机器人检测方法存在以下几个问题：（1）目前社交网络和用户自身隐私意识提高，导致依靠大量用户信息的社交机器人检测方法效果不佳。（2）现有恶意社交机器人检测方法只是将账户特征、推文特征融合，没有考虑用户之间交互行为，导致最终检测精度低。（3）对监测人员专业性要求较高，缺少自动化程度高的社交机器人检测系统提升检测效率。

根据上述存在问题，本文主要研究内容及创新点如下：

（1）针对用户隐私意识提高，导致依靠大量用户信息的社交机器人检测方法效果不佳的问题，本文提出基于账户特征的社交机器人检测方法。深入研究社交机器人与正常用户的账户特征，引入本福特定律筛选特征，减少特征维度；在此基础上将卷积神经网络中最后一层全连接层置换为零偏置全连接层，增加全连接层与Softmax之间的关联性，从而提高社交机器人检测精度。最后在不同数据集上验证了模型的有效性，在Cresci-2017数据集上本文提出的社交机器人检测方法准确率提高了2.2%，召回率提高了2%，MCC值达到了0.9675。

（2）针对目前恶意社交机器人检测没有考虑用户之间交互行为的社交关系，检测精度低的问题，在分析现有恶意社交机器人的社交行为后，本文提出基于图卷积神经网络的恶意社交机器人检测方法。通过账户的个性签名、配置信息等属性与社交网络相结合，利用多层图卷积捕捉各账户节点与邻居节点间复杂的交互，学习其中隐含信息。在此基础上利用社交账户的推文内容对分类结果进行修正，从而得到更准确的分类结果。仿真结果表明，在TwiBot-20数据集上本文构建的恶意社交机器人检测方法准确率提高了1.5%，F1分数达到了0.882，MCC值达到了0.7117。

（3）在上述方法基础上，本文设计和实现了一个社交机器人检测系统。社交机器人检测系统主要包括了数据采集模块、数据处理模块、社交媒体模块、社交机器人检测及可视化模块。通过社交机器人检测系统，提高社交机器人检测效率，缓解其对社交媒体和用户的危害。

Social media has become an indispensable part of People's Daily life and social communication. As a new form of Internet application, social media has a wide range of users and huge commercial value. However, there are a large number of social robots in social media, which seriously harm the security of user information and the healthy development of social media. In order to purify the cyberspace environment and reduce the threat of social robots, it is urgent to use effective detection technology to find social robots in social media, which is of great significance to the security of social media.

With the iterative evolution of social robots, the degree of "anthropomorphism" of social robots becomes higher and higher, which leads to the lower and lower detection efficiency. The existing detection methods of social robots have the following problems: (1) Currently, the awareness of privacy of social networks and users is increasing, which leads to the poor effect of the detection methods of social robots relying on a large amount of user information. (2) The existing detection methods of malicious social robots only integrate account features and tweet features without considering the implicit relationship between users' interaction behaviors, resulting in the final detection effect is mediocre. (3) Lack of highly automated social robot detection system to reduce labor consumption and mitigate its harm to social media and users.

According to the above problems, the main research contents and innovations of this thesis are as follows:

(1) To solve the problem that the social robot detection method relying on a large amount of user information is ineffective due to the increased privacy awareness of users, this thesis proposes a social robot detection method based on account characteristics. The account features of social robots and normal users are deeply studied, and Benford's law is introduced to screen the features and reduce the feature dimensions. On this basis, the last full connection layer in the convolutional neural network is replaced with a zero bias full connection layer to increase the correlation between the full connection layer and Softmax, thus improving the detection accuracy of the social robot. Finally, the effectiveness of the model was verified on different data sets. In the Cresci-2017 data set, the accuracy rate of the proposed social robot detection method increased by 2.2%, the recall rate increased by 2%, and the MCC value reached 0.9675.

(2) Aiming at the problem that the current detection of malicious social robots does not consider the implicit relationship between user interaction behaviors and has low detection accuracy, this thesis proposes a detection method of malicious social robots based on graph convolutional neural network after analyzing the social behaviors of existing malicious social robots. By combining account attributes such as individual signature and configuration information with social networks, complex interactions between account nodes and neighbor nodes are captured by multi-layer graph convolution to learn hidden information. On this basis, the classification results are corrected by using the Twitter content of social accounts, so as to obtain more accurate classification results. The final experimental results show that the accuracy rate of the malicious social robot detection method constructed in this thesis is increased by 1.5% on the TwiBot-20 data set, with the F1 score reaching 0.882 and the MCC value reaching 0.7117.

(3) Based on the above methods, a social robot detection system is designed and implemented in this thesis. The social robot detection system mainly includes data acquisition module, data processing module, social media module, social robot detection and visualization module. Through the social robot detection system, the detection efficiency of social robots can be improved and the harm to social media and users can be alleviated.