The rapid spread of fake news in Indonesia has significantly impacted social, economic, and political aspects, making reliable detection methods essential. This study introduces a fake news classification model based on Bi-LSTM architecture with IndoBERT embeddings and data augmentation using back translation. The dataset, collected from Detik.com and Turnbackhoax.id, underwent preprocessing and augmentation to enhance data diversity and model robustness. Experimental results reveal that the proposed approach achieves an impressive 99.58% accuracy, outperforming traditional models such as SVM (93.87%), Decision Tree (95.86%), and KNN (82.34%). Further comparisons show that the Bi-LSTM model without data augmentation achieves only 94.14% accuracy, while using Word2Vec embeddings with augmentation results in 97.26% accuracy. IndoBERT embeddings provide superior contextual representation compared to conventional Word2Vec, leading to higher precision, recall, and F1-score metrics. The combination of deep learning models, language-specific embeddings, and augmentation techniques proves highly effective in detecting fake news in Indonesian texts. These findings emphasize the importance of leveraging advanced NLP techniques to enhance classification performance, improve model generalization, and strengthen fake news detection efforts in Indonesia.
