TENSOR Detects Information Operations Users via Unsupervised Anomaly Detection

Information operations (IO) on social media pose a significant threat, yet their detection remains challenging due to evolving user behaviors and the limitations of existing methods. Supervised approaches struggle with the dynamic nature of IO, while current unsupervised techniques often rely on oversimplified assumptions about user coordination. To overcome these hurdles, researchers introduced TENSOR, an unsupervised anomaly detection framework that leverages multimodal data to identify IO users. TENSOR focuses on the unique temporal behavioral patterns, such as message posting activities, and the textual content of messages, recognizing that IO users constitute a small fraction with distinct characteristics. It employs a Temporal Point Process (TPP) to capture abnormal temporal behaviors, which are then adjusted by a novel evidence function. This function converts LLM responses, generated from user post timelines, into quantitative scores, enhancing the TPP's ability to detect IO users. Experimental results across five real-world IO datasets demonstrate TENSOR's superior performance compared to baseline methods, offering a more robust solution for combating online disinformation.