Multimodal Fake News Detection Using Distilbert And Xception Convolutional Neural Network

Abstract

The Rapid Proliferation Of Misinformation Across Digital Platforms Poses A Growing Threat To Public Discourse, Democratic Processes, And Societal Trust. Traditional Fake News Detection Approaches Rely Exclusively On Textual Analysis, Failing To Exploit The Deceptive Potential Of Associated Images. This Paper Proposes A Multimodal Deep Learning Framework That Combines DistilBERT For Efficient Semantic Text Analysis And The Xception Convolutional Neural Network For Visual Feature Extraction. DistilBERT, Derived From BERT Through Knowledge Distillation, Retains 97% Of BERT’s Language Understanding While Reducing Model Size By 40% And Improving Inference Speed. Xception Leverages Depthwise Separable Convolutions To Extract Discriminative Visual Patterns From News Images. Features From Both Modalities Are Concatenated Through A Fully Connected Fusion Layer And Classified Using Sigmoid Activation. The Proposed System Is Evaluated On The FakeNewsNet And GossipCop Datasets, Achieving 93.47% Accuracy With An F1-score Of 0.93, Outperforming Single-modality And Several Prior Multimodal Baselines. A Flask-based REST API Enables Real-time Deployment With Confidence Score Output.

Keywords