Deep Learning for Fake News Detection: A Comprehensive Survey of Adversarial Data Augmentation and Sentiment-Aware Paradigms

The rampant proliferation of misinformation and fake news across digital platforms poses unprecedented threats to social stability, public health, and democratic discourse. Recent advances in Natural Language Processing (NLP) and Deep Learning (DL) have enabled automated fake news detection; however, state-of-the-art models face two fundamental bottlenecks: severe class imbalance / data scarcity in real-world benchmarks, and the insufficient exploitation of implicit linguistic signals, notably affective and emotional cues. To address these limitations, recent literature has witnessed a surge in Generative Adversarial Network (GAN)-based data augmentation techniques and sentiment-aware representation learning. This paper presents a comprehensive narrative survey of deep learning paradigms in fake news detection with a dedicated focus on synthetic data generation and affective computing integration. We first introduce a structured taxonomy categorizing detection frameworks across textual, multimodal, and contextual dimensions. Subsequently, we critically examine adversarial augmentation pipelines (including Text-GANs, conditional GANs, and latent space perturbations) tailored for mitigating label skewness while preserving semantic fidelity. Furthermore, we investigate the integration mechanisms of sentiment polarity, emotion dynamics, and subjectivity features into deep architectures (such as Bi-LSTMs, Transformers, and Graph Neural Networks). Finally, we systematically benchmark open datasets and evaluation metrics, synthesize key empirical findings, and outline critical open challenges—such as adversarial robustness, cross-domain generalization, and explainability—to guide future research trajectories.