Multimodal Fake News Detection Using Deep Learning Techniques

In the current digital world, fake news is widely disseminated on social media sites. An unverified news is often disseminated through text and images. Before sharing the news with the public sector, it is crucial to confirm it. Earlier, unimodal fake news detection (FND) techniques are not much effective at detecting false news as they extract and classify textual and visual features separately. In order to identify false news, multimodal fake news detection models have been proposed. These models use textual and visual data to identify fake news. The textual features are extracted using the Bidirectional Encoder Representations from Transformers (BERT) model, which keeps the semantic links between words. The Capsule Neural Network (CapsNet) derives extremely informative visual feature from an image. Utilizing these attributes produces a richer data representation that makes it simpler to determine if the news is authentic or false. The most of the model only examines dataset of fake news in english to determine if it is true or fake, but CB-Fake approach is expanded to incorporate datasets of fake news in other widely used languages, such as Tamil. Robustly Optimized BERT Pre-training Approach (Roberta) is used to classify multilingual textual features. The Gossipcop dataset is utilized to evaluate the performance of the approach. The classification accuracy of proposed model of BERT and Roberta are 72% and 78% respectively.