Contrastive Learning-Driven Fake News Detection: Preserving Semantics, Unveiling Distortions.

Abstract

The complex interweaving of authentic and fake news on social networks poses formidable challenges to modern information management. Existing fake news detection methods heavily rely on extensive labeled data and auxiliary information, such as propagation structures, facing critical issues of data scarcity and semantic integrity destruction in practical applications. Traditional data augmentation methods, due to their fixed rule-based transformation patterns, fail to effectively simulate the complexity and irregularity of information propagation in real social networks. To address these fundamental limitations, this article proposes a contrastive learning-driven fake news detection (CLFD) framework that breaks through existing technical bottlenecks through innovative distortion-reversion dual-view manipulation mechanisms and distortion-aware contrastive learning methods. The core innovation of CLFD lies in employing learnable neural networks to precisely simulate nonlinear information transformation processes, generating stylistically diverse contrastive views while preserving semantic core integrity, fundamentally solving the critical problem of semantic destruction in traditional methods. More importantly, our method achieves efficient detection using only textual content, requiring no additional information such as propagation structures or social network topologies, demonstrating outstanding universality and portability. Through dynamic view generation and multiobjective joint optimization strategies, CLFD significantly enhances the model's capability to capture deceptive features in fake news. Extensive experiments on multiple benchmark datasets demonstrate that our framework significantly outperforms existing state-of-the-art methods in detection accuracy, robustness, and generalization capability.