Successful disinformation campaigns depend on the availability of fake social media profiles used for coordinated inauthentic behavior with networks of false accounts including bots, trolls, and sockpuppets. This study presents a scalable and unsupervised framework to identify visual elements in user profiles strategically exploited in nearly 60 influence operations, including camera angle, photo composition, gender, and race, but also more context-dependent categories like sensuality and emotion. We leverage Google’s Teachable Machine and the DeepFace Library to classify fake user accounts in the Twitter Moderation Research Consortium database, a large repository of social media accounts linked to foreign influence operations. We discuss the performance of these classifiers against manually coded data and their applicability in large-scale data analysis. The proposed framework demonstrates promising results for the identification of fake online profiles used in influence operations and by the cottage industry specialized in crafting desirable online personas.

中文翻译：

---

四万个虚假 Twitter 个人资料：社交媒体宣传视觉分析的计算框架


成功的虚假信息活动取决于虚假社交媒体资料的可用性，这些资料用于与机器人、巨魔和马甲等虚假帐户网络协调不真实的行为。这项研究提出了一个可扩展且无监督的框架，用于识别在近 60 个影响操作中战略性利用的用户配置文件中的视觉元素，包括相机角度、照片构图、性别和种族，但也包括更多依赖于上下文的类别，如性感和情感。我们利用 Google 的 Teachable Machine 和 DeepFace Library 对 Twitter Moderation Research Consortium 数据库中的虚假用户帐户进行分类，该数据库是与外国影响力操作相关的社交媒体帐户的大型存储库。我们讨论这些分类器针对手动编码数据的性能及其在大规模数据分析中的适用性。所提出的框架在识别影响力操作和专门制作理想在线角色的家庭手工业中使用的虚假在线个人资料方面展示了有希望的结果。