Beyond Text: Detecting Image Propaganda on Online Social Networks

Abstract

The rapid expansion of social media has notably transformed political communication, with politicians and activists increasingly adopting multimedia formats to disseminate their ideologies and policy proposals. This transformation poses a significant risk of propaganda through coordinated campaigns that leverage template-based imagery to spread political messages. To tackle this challenge, our research focuses on developing a detection framework for identifying political images crafted from similar templates, which are a common tool in such propaganda efforts. During a national referendum held in 2021 in Taiwan, we collected visual content from various social networks and implemented a hybrid approach that combines object recognition, textual analysis, and pixel-level information. This methodology is specifically designed to detect patterns and similarities within propaganda images, enabling us to trace and analyze the potentially manipulative content. Our hybrid feature combination technique has demonstrated superior performance compared to several established baseline methods in identifying template-based images. This advancement in detection technology not only enhances the efficiency of researchers studying political communication but also serves as a crucial tool in uncovering and understanding the mechanisms behind potential political propaganda and coordinated efforts to shape public opinion on social media platforms.