在线社交网络文本内容对抗技术  

Adversarial Technology of Text Content on Online Social Networks

作　　者：刘晓明[1] 张兆晗 杨晨阳 张宇辰 沈超[1] 周亚东[1] 管晓宏[1] LIU Xiao-Ming;ZHANG Zhao-Han;YANG Chen-Yang;ZHANG Yu-Chen;SHEN Chao;ZHOU Ya-Dong;GUAN Xiao-Hong(Faculty of Electronic and Information Engineering,Xi′an Jiaotong University,Xi′an 710049)

机构地区：[1]西安交通大学电子与信息学部,西安710049

出　　处：《计算机学报》2022年第8期1571-1597,共27页Chinese Journal of Computers

基　　金：国家自然科学基金(61902308,62103323,61822309,61773310,U1736205,U1766215);博士后创新人才支持计划基金(BX20190275,BX20200270);博士后面上基金(2019M663723,2021M692565);西安交通大学基本科研经费(xjh032021058,xxj022019016,xtr022019002)资助.

摘　　要：在线社交网络内容对抗技术是人工智能领域与网络空间安全领域的一个新兴研究方向.它是指人们基于特定任务,在社交网络受众广泛、内容数量庞大、内容质量参差、内容真伪难辨的环境下,利用新兴的大数据驱动的人工智能方法,自动完成在线社交网络中针对特定主题与群体的对抗内容发现、生成与投送,进而实现社交平台异常信息的检测与反制,以达到维护网络空间安全的目的.虽然在线社交网络对抗技术属于一个新概念,鲜有相关工作,但是已有的机器学习方法可以被应用到该领域,通过特征提取、文本模式解析、文本内容编码与重建、目标优化等技术对社交文本大数据进行解析与内容表示,解决网络空间中的文本内容安全问题,实现对社交网络文本环境的净化.此外,在社交网络文本内容对抗的过程中,对抗双方的策略可作为反馈信息,使对抗模型不断进行更新和优化,最终达到完善模型的目的.基于以上攻防对抗思想,本文着重从文本内容生成与检测两方面对在线社交网络对抗进行阐述.首先,本文介绍了有关在线社交网络文本对抗技术的相关基础知识.其次,针对社交网络文本内容检测方法,本文从基于零次分类器的模型、基于机器特征的模型、基于预训练语言模型的方法、基于人机协作的模型、基于能量基础的模型5个角度进行详细介绍.为了方便读者针对不同的应用场景选择合适的模型,本文对不同检测模型的适用场景以及模型优劣进行了对比总结.针对社交网络文本生成方法,本文对基于对抗生成网络的文本生成模型、可控文本生成模型、长文本生成、文本质量评价4方面进行了综述.此外,为了方便读者对模型的有效性进行验证,本文对相关数据进行了系统性总结.最后,本文总结了在线社交网络对抗技术未来的重要研究方向与挑战.With prosperous development of social network and intensive research on natural language processing technology,adversarial technology of text content on online social networks becomes an emerging research direction in the field of artificial intelligence and cyberspace security.Adversarial technology of text content on online social networks refers to that in social network environment where there exists numerous users,huge amount of text contents,uneven quality of texts,ambiguous credibility of information,people employ the latest artificial intelligence methods to accomplish specific tasks such as discovery of machine-generated text,coherent and controllable text generation and adversarial content delivery for specific topics and targeted groups in online social networks automatically and precisely to filter the content in online social networks based on analysis of generation strategies,appropriate representation pattern of aimed groups,which could realize the detection and countermeasures of social platform abnormal information attack,increase the information credibility in social network,improve the quality of daily news received by social network users,enhance public trust on social media and protect cyberspace from information bombing and misleading information by reliable and economic means.Although online social network adversarial technology is a new concept with few related work,existing machine learning methods can be applied to this field by applying advanced technologies such as feature extraction,document parsing,encoding and reconstruction of text content,objective optimization on large scale social network text content dataset to solve the practical problem and clear the Internet environment.Besides,during the adversarial process of text content on online social networks,the strategies of the opposing parties can be used as feedback information to each other,so that the adversarial model is continuously updated and optimized,and finally the goal of perfecting the model is achieved.Based on the ad

关 键 词：社交网络对抗 人工智能 网络空间安全 文本内容自动生成 机器生成内容检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]