Multi-Model Fake News Detection based on Concatenation of Visual Latent Features

Online Social media for news consumption is a double-edged sword. If we ponder on the positives outcomes for this, it includes easy access, negligible cost, smart categorization and out reach to the very customer in seconds. But, as every coin has two sides and when we flip side of this, a series of issues come up which need immediate attention and most important among them is spreading of fake news. This has become a serious threat for the governments of countries to keep their harmony intact, keep faith of public in democracy and justice and sustenance of public trust. Therefore fake news detection, especially in social media platform has become an emerging research topic that is attracting tremendous attention. Current set of detection algorithms are specially showing their inability to learn the shared representation of texts and visuals combined (popularly known as multimodal) information. Therefore, we present a variational auto encoder based framework, which consists of three major components encoder, decoder and fake news detector. It utilize the concatenation of visual latent features from three popular CNN architecture (VGG19, ResNet50, InceptionV3) combined with textual information to detect fake news with the help of binary classifier. We conducted the experiment on publically available Twitter dataset. The experimental result shows that out model improves state of the art method by the margin of $\sim$2% in accuracy and $\sim$3% in F1-score.