Empirical Framework for Automatic Detection of Neural and Human Authored Fake News

The evolution of digitization and technology has increased the amount of data shared online, inevitably leading to the spread of false information. The issue has skyrocketed as COVID-19 spread across the globe and carried with it a sea-full of fake news. Recent technological advances, such as automatic text generators, have exacerbated the problem by interpolating synthetic text into the fake news. Current fake news detection approaches do not take into account the validity of synthetic news (machine-generated news), thus classifying all machine-generated material as fake. In this paper, the first-ever synthetic news classification model using CT-BERT is implemented, and a framework is proposed that not only distinguishes between human-authored and machine-generated text, but also considers the veracity of text to detect fake news. Moreover, a novel COVID-19 based synthetically generated dataset has been introduced by fusing synthetic text generated by GPT-2 and Grover model. Further, whether the GPT-2 and Grover models are vulnerable to adversarial attacks or not has been investigated. An accuracy of 98.2% and 92.4% respectively in the fake news classification of human-authored and machine-generated text have been achieved.