Risk Traceability of False Content Generation Enabling by Machine Learning Technology in the Intelligent Media Era

Abstract

This paper employs a Text-CNN network to extract textual and image features from multimodal fake media content, maps different media features into a joint space, and achieves comprehensive feature extraction of multimodal media content through concatenation and embedding. We propose MTTV, a media false content detection method based on a multimodal Transformer and designed for multidimensional feature information. It embeds the extracted multimodal information into the detection network and uses a multi-layer Transformer encoder to classify content as true or false. Results indicate that the most important multimodal features, in descending order, are follower count, like count, user reputation score, registration duration, and post count, with importance values of 0.312, 0.305, 0.300, 0.271, and 0.250, respectively. The model proposed in this paper achieves an accuracy rate of 0.922 for media fake content detection, the highest among all compared models. Based on the features identified for fake content detection, mitigating the risk of media fake content generation requires the joint participation of multiple stakeholders, including media professionals, commercial platforms, and the public.