Multimodal fake news detection on social media: a survey of deep learning techniques

The escalation of false information related to the massive use of social media has become a challenging problem, and significant is the effort of the research community in providing effective solutions to detecting it. Fake news are spreading for decades, but with the rise of social media, the nature of misinformation has evolved from text-based modality to visual modalities, such as images, audio, and video. Therefore, the identification of media-rich fake news requires an approach that exploits and effectively combines the information acquired from different multimodal categories. Multimodality is a key approach to improving fake news detection, but effective solutions supporting it are still poorly explored. More specifically, many different works exist that investigate if a text, an image, or a video is fake or not, but effective research on a real multimodal setting, ‘fusing’ the different modalities with their different structure and dimension is still an open problem. The paper is a focused survey concerning a very specific topic which is the use of deep learning (DL) methods for multimodal fake news detection on social media. The survey provides, for each work surveyed, a description of some relevant features such as the DL method used, the type of analysed data, and the fusion strategy adopted. The paper also highlights the main limitations of the current state of the art and draws some future directions to address open questions and challenges, including explainability and effective cross-domain fake news detection strategies.