Deepfake News Detection Using Bidirectional Encoder Representations from Transformers

Deepfake news detection is the act of detecting and recognizing altered or synthetic media content, especially when it appears in news, videos or other journalistic content. Deepfakes are modified or AI-generated media, ranging from images or videos that are often intended to deceive viewers by pretending that fictitious information or events are authentic. Social media has come out as an enormously efficient tool for disseminating news quickly and widely. Its rapid data exchange and ease of use have fostered to its extensive use. Users from diverse range of age groups, genders and socioeconomic backgrounds actively engage in social media discussions. On the other hand, the widespread circulation of fake news is a major disadvantage because people readily share and consume information without verifying it for veracity. As a result, learning the methods to check the accuracy of news becomes a necessity. In this study, a transform-based Bidirectional Encoder Representations from Transformers technique is used. The Bidirectional Encoder Representations from Transformers language model is an open source machine learning framework for natural language processing (NLP). BERT aims to aid computers in understanding ambiguous text by providing context through surrounding content. The Bidirectional Encoder Representations from Transformers, refers to a deep learning model in which each input and output element is coupled to another and the weights between them are dynamically determined based on the ratio of importance of different words and relevance to the current word being processed. In deep neural network architecture, BERT offers a special framework for detecting fake news. The three components of a neural network are layers, neurons and weights. Layers of neurons are weighted to adjust for the effects of various inputs and biases. The complexity and design of this architecture is closely relevant to the task at hand. In this scenario, a base model is created and a pre-trained model is acquired. Once the layers are frozen, a new layer of the information sheet is taught. Subsequently, the model is refined and the outcomes are anticipated.