Roberta-LightGBM: A hybrid model of deep fake detection with pre-trained and binary classification

Abstract

In May 2023, a fake image of an explosion near the Pentagon gained widespread social media traction. It dragged down US markets momentarily, perhaps marking the first time an artificial intelligence (AI)-generated image has affected the market. The fictitious image originally surfaced on Facebook and showed a large column of smoke that a Facebook user said was close to the US military headquarters in Virginia. In this research, we proposed Roberta by combining lightGBM to construct the Roberta-LightGBM technique framework. This paper aims to reduce tampered fake content in media with good accuracy and faster mechanisms by designing these two approaches to detecting fake content using a natural language model and a machine learning algorithm combined to develop the proposed work. Roberta`s NLP model helps us train large datasets in minimum time, compared to traditional techniques like the BERT technique, which requires ten times larger datasets to be trained in a wide range of applications. LightGBM was used to identify the solution of a machine learning algorithm using a decision tree to involve binary classification to predict whether the retrieved data was real or fake. It improved the faster training speed in handling large datasets with high accuracy; memory usage was reduced, resulting in better accuracy. As a result of the analysis, the proposed framework achieves the goal of this research when compared to alternative techniques such as the XGBoost technique, the Roberta-LightGBM technique gives 95.36% accuracy, the overall computational time is 4.4 seconds, and the implementation of Roberta to get 92.17% efficiency is shown experimentally in this paper.