Generative Adversarial Networks for Mitigating Bias in Disinformation

The increasing use of social platforms has led to the rampant spread of disinformation in a number of domains and hampered the credibility of these sources. Disinformation is a type of information disorder wherein fake information is generated with the intent to harm. Machine learning and deep learning can be used to detect such auto-generated information. The conventional method includes gathering vast volumes of text data in terms of both true and fake news stories, headlines, or any other sources and then labeling the data. This leaves us with the issues of human bias within these existing systems, the quality of the aggregated data from different sources and their accurate labeling. In this paper, we propose and demonstrate a textual deep fake detection and generation framework based on Generative Adversarial Network (GAN). The methods are demonstrated on three separate domains namely, political data, sports-related data, and medical data pertaining to COVID 19. Cosine similarity-based evaluation metric is employed for analyzing the classification accuracy of the discriminator and the performance of the Generator in generating near-true fake articles which are very similar in the style and context to the real datasets. The contribution of this work is threefold; firstly, we have demonstrated the generator network which generates the synthetic textual deep-fakes which are extremely similar to the actual instances of disinformation. Secondly, the development of a discriminator network that detects the system-generated data with the real data and lastly, the composite GAN model which can be used in various applications such as automated labeling and annotation is shown. Additionally, this approach can help in generating balanced datasets required for machine learning tasks and alleviate the problem of bias in AI systems, especially for disinformation detection tasks where the disinformation class instances are typically lesser as compared to the true information instances.