ABERT: Adapting BERT model for efficient detection of human and AI-generated fake news

The proliferation of fake news in digital media poses a significant challenge to the dissemination of accurate information. Transfer learning, particularly with pre-trained language models (PLMs) like BERT, has demonstrated exceptional performance in natural language processing (NLP) tasks. However, the computational expense of fine-tuning the entire model for domain-specific tasks remains a limitation. In this study, we propose a novel approach, Adapt-BERT (ABERT), for the detection of both human and artificial intelligence (AI)-generated fake news. ABERT includes parameter-efficient adapter that enables efficient detection. By freezing the pre-trained BERT network and incorporating lightweight adapter, ABERT achieves comparable performance to fully fine-tuned BERT while reducing the number of trainable parameters by approximately 67.7%. ABERT strikes a balance between performance and computational efficiency, offering a scalable solution to combat the dissemination of fake news in digital media. Experimental evaluations on diverse datasets showcase the effectiveness of the proposed parameter-efficient approach in achieving comparable performance to state-of-the-art (SOTA) methods in the task of fake news detection (FND). • Propose ABERT: a lightweight BERT with adapters and fusion for efficient FND tasks. • ABERT is significantly more efficient than fully fine-tuned BERT model. • ABERT matches BERT performance while reducing trainable parameters by 67.7%. • Tested on human- and AI-generated news to demonstrate the method’s effectiveness. • Ablation study explores adapter injection methods to enhance ABERT’s efficiency.