Enhanced Multimodal Fake News Detection with Optimal Feature Fusion and Modified Bi-LSTM Architecture

Abstract

Numerous enhancements have been made to the mobile internet, which leads to an increase the people’s attention to posting more multi-modal posts among the social media platforms. Hence, this paper aims to design a multimodal fake news detection model with enhanced deep learning architecture. Initially, the multi-modal information including images and text are gathered from real-time social media. Then, pre-processing of both images and text is carried out. Further, the text features are extracted using Word2vector and glove embedding techniques. Then, the optimal features from both image features and text features are attained using the Adaptive Water Strider Algorithm (A-WSA). The achieved optimal features are forwarded to the new feature fusion concept based on the weight factor that is optimized using the same A-WSA. Finally, the fused features are forwarded to the fake news classification stage with the help of “Optimized-Bidirectional Long Short-Term Memory (O-BiLSTM),” where the hyperparameter of BiLSTM is optimized through similar A-WSA. Throughout the result analysis, the accuracy rate of the designed A-WSA-BiLSTM method is attained at 96.51%. The experimental result demonstrates that the proposed model is effective in using multi-modal data in automated fake news classification.