Disinformation at scale functions as an attack on the integrity of online information systems and on the trust users place in them, making automated detection a problem of practical interest to security and trust-and-safety practitioners as well as NLP researchers. We present a comparative study of three classical machine learning approaches to fake news detection: Naive Bayes, Logistic Regression, and Stochastic Gradient Descent (SGD), trained on a labeled corpus of real and fake news articles under two text-vectorization schemes (count-based and TF-IDF). Logistic Regression with TF-IDF features achieved the highest test-set accuracy at 93.9%, followed by SGD at 91.9%, with Naive Bayes trailing at 82.0–89.1% depending on the vectorizer. A confusion-matrix analysis reveals that headline accuracy alone is misleading for this task: the deployed Naive Bayes pipeline's lower score is driven almost entirely by weak recall on the FAKE class, meaning it disproportionately lets fake articles through undetected rather than failing symmetrically on both classes. We discuss why this gap emerges from Naive Bayes' independence assumption and surface it as a practical caution for anyone selecting a classical model for misinformation detection on accuracy alone. As a deployment vehicle for the trained models, we built a Flask-based web application (truDefender) that lets users classify a news article either by pasting its text or supplying a source URL, demonstrating that classical, computationally cheap models remain viable for real-time fake news triage. Note: this work was originally completed as a B.E. capstone project at MVJ College of Engineering during the 2023–2024 academic year. This manuscript is a revised and extended write-up of that project, prepared for preprint publication in 2026.
R et al. (Fri,) studied this question.