ENHANCING MULTIMODAL FAKE NEWS DETECTION: OPTIMIZED HIERARCHICAL DEEP LEARNING VIA ADAPTIVE EVOLUTIONARY ALGORITHMS

Abstract

The rapid spread of multimodal fake news on social media, combining deceptive text, manipulated images, and misleading metadata, poses a severe threat to public discourse and democratic processes. Conventional single-modality detectors achieve limited accuracy (≤78%), as they fail to capture cross-modal interactions critical for identifying sophisticated misinformation.

This paper proposes the Deep Learning with Evolutionary Computing Approach (DLECA), a novel framework that simultaneously compresses and optimizes hierarchical deep learning models (HDLMs) for multimodal fake news detection. DLECA employs an enhanced genetic algorithm featuring: (1) a multi-objective fitness function balancing accuracy and parameter reduction; (2) dynamic crossover that adapts probability based on population fitness divergence; (3) adaptive mutation synchronized with crossover evolution; and (4) automated architecture search replacing manual hyperparameter tuning.

Evaluated on benchmark multimodal datasets, DLECA achieves up to 97.86% model compression (reducing parameters from millions to ~35K) while improving accuracy by 0.34% over baseline HDLMs. A lightweight variant delivers 96.24% compression with a 0.23% accuracy gain. Comparative analysis demonstrates DLECA’s superiority over Random Walk and Bayesian Optimization in convergence speed, final accuracy, and resource efficiency.

By automating complex architectural decisions, DLECA enables deployment on resource-constrained devices, advancing real-time fake news detection. This work establishes a scalable paradigm for evolutionary deep learning optimization across multimodal classification tasks.