A Systematic Review of Deep Learning Approaches for Fake News Detection

Abstract

Distinguishing real news from fake news on online platforms remains a critical research challenge. This systematic literature review (SLR) evaluates hybrid and non-hybrid deep learning (DL) approaches for fake news detection across ISOT, WELFake, and FA-KES datasets, comparing their mean accuracy and model performance. PRISMA guidelines were followed to search databases such as Google Scholar, Scopus, IEEE Xplore, and MDPI using search terms such as "fake news detection", "deep learning", and "hybrid models" for articles published between 2018 and 2025. Out of 154 articles surveyed, 34 duplicates were removed, 70 were excluded for irrelevance, and 50 were included. Addressing three research questions shows that hybrid models like Convolutional Neural Networks and Long Short-Term Memory (CNN-LSTM) consistently outperform non-hybrid models (e.g., CNN) by 0.16% – 8.6% accuracy across the three datasets (mean accuracy: hybrid 85.46%; non-hybrid 82.46%). Combining the right models in a hybrid approach improves performance, yielding more reliable results. Limitations include dataset biases. Future research could explore multimodal datasets and BERT base models to improve robustness.