Fake News Generation and Detection: Adversarial use of Generative AI in Text Synthesis

Abstract

The extensive growth of online media platforms, as well as the extensive use of generative artificial intelligence, has made the production and distribution of fake news particularly more active. Models of languages and vision have now been advanced to create a high level of reality in text, images and videos that become even harder to detect any misinformation by means of the old standard verification channels. The paper is a detailed analysis of fake news and its main premises, the core issues, and the role of generative AI in the development and detection of fake news that keeps changing. The paper examines uni-modal deepfake detection models in text, image and video modalities, and cross-modal consistency analysis, which utilizes semantic consistency between heterogeneous data. In addition, the study contrasts traditional machine learning approaches with more modern ones, detailing the benefits and downsides of each, as well as their suitability for large-scale and multilingual tasks. A number of popular benchmark datasets are presented to give an idea of the experimental evaluation practice. Additionally, stability metrics such as recall (REC), accuracy (ACC), precision (PRE), F1-score (F1), and ROC-AUC are assessed to ascertain the efficacy of detection. Find that deep learning models are more generalizable and perform better overall, but they are still vulnerable to adversarial manipulation and new generative methods.