A Novel Enhanced CNN Framework for Automated Fake News Detection

Abstract

Now a days a smartphone can publish a news story and reach millions of readers within hours. While the democratisation of information has real advantages, it has also created conditions for misinformation to spread at a scale and speed that earlier generations simply never faced. Fake news deliberately fabricated or distorted content dressed up to look like legitimate journalism now threatens democratic processes, public health, and basic social trust in ways that are difficult to overstate. Social media platforms and news aggregators amplify this problem further, because their algorithms are optimised for engagement, and emotionally provocative falsehoods tend to be far more engaging than carefully hedged truths. The work unfolded across five broad stages: assembling a large labelled dataset from multiple benchmark sources, running a thorough exploratory analysis to identify what actually distinguishes fake news from real journalism at a linguistic and structural level, comparing how well different classification algorithms handle this task, designing and validating a new hybrid architecture I call HybridFND, and finally thinking through what a real-world deployment of such a system would look like. Ablation experiments confirmed that both branches genuinely earn their place in the model removing either one hurts performance, with the BiLSTM on its own. The toughest cases were satirical articles (which share surface features with real fake news), hybridised content that mixes genuine facts with fabricated claims, and stories about rapidly evolving events where the truth was still being established at the time of publication.