Automatic Fake News and Deepfake Detection Using Machine Learning and Deep Learning Models

The spread of misinformation, both in text and multimedia formats (deepfakes), poses a serious threat to public trust and social stability. This paper is a comprehensive empirical study of machine learning (ML), deep learning (DL), and cutting‐edge transformer models to detect multimodal misinformation. To address generalizability and dataset bias issues, experiments were conducted on the ISOT and real‐world FakeNewsNet datasets for text, and empirical validation was conducted on Celeb‐DF for deepfakes. Furthermore, external benchmarking was conducted on the LIAR and FaceForensics++ datasets via leave‐one‐dataset‐out validation to quantify domain adaptation capability. All experiments followed strict data leakage prevention strategies and standardized preprocessing pipelines. In the case of textual fake news, classical ML models (e.g., random forest) performed well when given structured data in ISOT (99.7%) but experienced significant performance degradation on noisy, real‐world data. On the other hand, transformer‐based models such as RoBERTa showed better cross‐domain generalization (92.1% accuracy on FakeNewsNet). In the visual misinformation scenario, the Vision Transformer (ViT) achieved higher accuracy (98.4%) than the standard CNN (XceptionNet) on Celeb‐DF, due to its global self‐attention mechanism, which effectively identifies subtle facial artifacts. To jointly exploit these modalities, early and late fusion architectures with cross‐modal attention were introduced for end‐to‐end detection. The evaluation framework was expanded with robust metrics (ROC–AUC, PR‐AUC, MCC), mean ± standard deviation, Cohen’s d effect sizes, and confidence intervals, alongside McNemar’s statistical significance testing. Additionally, the study provides comprehensive ablation studies, model calibration analysis, adversarial robustness evaluation, explainability via SHAP and Grad‐CAM, and an expanded computational efficiency analysis (including FLOPs and parameter counts). The results show a key trade‐off: The greatest accuracy and contextual understanding come with the transformer, which requires a lot of memory and inference time, but classical ML models are quite viable in real‐time and resource‐constrained deployments. This study provides a clear, reproducible, and transparent framework for developing robust, modality‐aware misinformation defense systems.