<![CDATA[The spread of fake news through digital media platforms has become a serious problem affecting social, political, and public health conditions in Indonesia. The speed of information dissemination has exceeded the capacity of existing manual verification mechanisms. This study aims to develop an automated fake news detection system for Indonesian-language news using a fine-tuning approach on two BERT variants, namely IndoBERT and multilingual BERT (mBERT), with a multi-field input strategy that combines news headlines and narratives into a single sequence using the [SEP] separator token. The dataset consists of 22,400 balanced samples collected from TurnBackHoax.id, Antaranews, Kompas, and Detik, and is divided into training, validation, and testing sets using a 70:15:15 ratio. Both models were trained using identical hyperparameter configurations on the Kaggle platform with dual T4 GPU acceleration. Evaluation results on the test set show that mBERT achieved an accuracy of 99.79% and an F1-score of 0.9979, slightly outperforming IndoBERT which achieved an accuracy of 99.67% and an F1-score of 0.9967. The system was implemented as a web application using FastAPI and HTML/JavaScript, featuring real-time prediction, automated article scraping from seven news websites, and LIME-based explainability visualization integrated directly into the user interface. Black-box testing achieved a pass rate of 95.5%, while white-box testing achieved 100%, confirming that the system functions according to the designed specifications. ]]>
Copyrights © 2026
