Hierarchical Deep Learning Model Optimization Using Enhanced Evolutionary-based Approach for Fake News Detection

Abstract

Multimodal fake information on social media is a growing concern worldwide. Existing deep learning-based solutions typically involve designing hierarchical models that capture relevant features from each modality, which are then fused for final classification. However, these models are often complex, with numerous trainable parameters, making them resource-intensive. This work introduces the Deep Learning Model with Evolutionary Computing Approach (DLECA), a novel method for compressing and optimizing hierarchical deep learning models (HDLM). It employs an enhanced genetic algorithm (GA) with a unique fitness function, dynamic crossover, and adaptive mutation strategies to achieve model compression, maintain accuracy, and balance exploration and exploitation during evolution. In comparison to manually designed HDLM, the proposed approach achieves up to 97.86% model compression with a 0.34% accuracy improvement, while a variant achieves 96.24% compression with a 0.23% accuracy improvement. Comparative analysis shows that DLECA outperforms Random Walk and Bayesian Optimization in multimodal fake news detection, offering a more efficient and accurate solution.