Abstract
The utilisation of fake news as a tool of soft power is regarded as a strategic component of contemporary warfare, posing an escalating threat to societal and political stability as well as public trust. Deep Learning models achieve superior performance; yet, they require substantial computational resources and possess interpretative restrictions, which hinder their practical application in real-time scenarios and impose constraints on resource configurations.
This study evaluates a traditional machine-learning framework based on linear and ensemble classifiers, including Linear Support Vector Classification (LinearSVC), Logistic Regression, Multinomial Naïve Bayes, and Random Forest. The framework employs a fixed, explicitly specified TF–IDF vectorization configuration and a consistent preprocessing pipeline. The system was assessed using the WELFake dataset, which contains over 70,000 labelled news stories, employing accuracy, precision, recall, F1-score, and ROC-AUC as evaluation metrics. The findings demonstrate the superiority of LinearSVC, attaining the maximum accuracy of 95.65%, providing balanced performance metrics, and surpassing all other models. This study contributes by providing an interpretable, scalable, and domain-agnostic solution through the development of a practical fake news detection system designed to combat digital disinformation in real-world contexts.