DEFT-Net: Explainable Deepfake Text Detection for Combating Information Disorder in the Age of Generative AI

The rapid evolution of Generative AI and large language models has drastically increased the risk of information disorder across digital communication platforms. The synthetically created messages with a deepfake aiming to mislead, manipulate the opinion of the masses, or disinformation is a serious danger to the honesty and credibility of online communication. Addressing this growing challenge requires an effective and transparent detection framework that can operate effectively at scale. This study introduces DEFT-Net (Deepfake Explainable FastText Tri-Convolutional Network), a practical and explainable deep learning architecture specifically designed for deepfake text detection on social media platforms. DEFT-Net leverages a multi-scale convolutional neural network (CNN) architecture combined with FastText embeddings to capture both syntactic and semantic patterns characteristic of synthetic content. The model was trained and evaluated using the TweepFake dataset, a rich collection of real and deepfake tweets, enabling the development of an efficient classifier capable of distinguishing authentic from fabricated text. To ensure transparency and foster trust in AI-based decision-making, this work integrates Explainable AI (XAI) techniques to provide interpretability of the model’s predictions. Experimental results demonstrate that DEFT-Net achieves high detection performance attaining 93% accuracy, 92% precision, 95% recall, and a 93% F1-score, outperforming several evaluated baseline models on the TweepFake dataset using FastText, subword embeddings, BERT embeddings, and GloVe. The proposed framework can further aid in the overall combating of information disorder by providing a practical, explainable, and computationally efficient solution for deepfake text detection, which may have potential uses in digital forensics, online safety, and safeguarding information integrity.