M3DUSA: A Modular Multi-Modal Deep fUSion Architecture for fake news detection on social media

Abstract Modern social networks facilitate the rapid global dissemination of news. However, much of this content is often unverified or shared based on personal opinions and beliefs, leading to widespread misinformation, erosion of public trust, and potential social and political instability. In this evolving landscape, the early detection of fake news has become a pressing challenge. To address this issue, multi-modal approaches have gained increasing attention due to their ability to integrate diverse sources of information, including textual content, images, and network structures, to capture richer contextual cues and deeper semantic relationships, improving detection performance. Despite the growing interest in this field, there remains no clear consensus on the most effective fusion strategy for integrating these heterogeneous data sources. In this work, we introduce M3DUSA $$^1$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mrow/> <mml:mn>1</mml:mn> </mml:mmultiscripts> </mml:math> , a modular multi-modal framework designed to effectively detect malicious content by leveraging the complementary strengths of multiple modalities. By systematically exploring the interplay between different fusion techniques, our framework adapts to various fake news detection scenarios, offering flexibility and robustness in identifying deceptive content across diverse contexts. To evaluate the effectiveness of M3DUSA , we conducted extensive experiments on two real-world datasets. The results demonstrate that our framework outperforms existing methods, highlighting its superior ability to exploit multi-modal information for improved fake news detection. Our findings confirm that while both early and late fusion strategies can effectively capture complementary signals across modalities, their advantages vary depending on the detection objectives and dataset characteristics.