Fake News Detection Using Deep Learning with Continual Learning

Abstract

Misinformation propagation across online platforms continues to pose serious risks to informed public discourse and media credibility. To address this, we design and evaluate a fully integrated fake news detection pipeline built upon the FakeNewsNet benchmark, drawing from both PolitiFact and Buz-zFeed corpora. This work combines conventional machine learning specifically TF-IDF-driven Logistic Regression with neural architectures including BiLSTM with attention and BERT, further augmented by continual learning mechanisms. Rigorous comparative evaluation using accuracy, F1-score, and ROC-AUC reveals that the TF-IDF with Logistic Regression configuration attains a macro F1-score of 0.607 alongside 92.3% accuracy on PolitiFact, whereas BERT leads with the highest ROC-AUC of 0.73. Beyond text, the system handles image-to-text conversion to support classification of screenshots and meme-based content. The continual learning component first trains on PolitiFact and subsequently adapts to BuzzFeed, emulating real-world incremental domain shifts. A production-grade REST API paired with an intuitive web interface completes the deployment.