International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 12 | Dec 2025
p-ISSN: 2395-0072
www.irjet.net
A Comprehensive Survey of Deep Learning Techniques for Fake News Detection: Emerging Trends and Future Directions Ajay S Tippannavar Ajay S Tippannavar, Department of Artificial Intelligence and Data Science, S J C Institute of Technology, Chikaballapur, India, -------------------------------------------------------------------------***-----------------------------------------------------------------------deceive or mislead. The consequences of such ABSTRACT - The exponential proliferation of digital misinformation ("fake news") on social media platforms presents a critical challenge to information integrity and democratic stability. The high velocity and volume of usergenerated content render manual verification methods unscalable. This research paper presents a comprehensive, systematic survey of Deep Learning (DL) methodologies employed for automated fake news detection, covering the period from 2022 to 2025. We critically analyse the architectural evolution from Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to state-of-the-art Transformer-based models like BERT and Roberta. Furthermore, we examine the emerging threat of Large Language Model (LLM) generated disinformation. Our comparative analysis demonstrates that while Transformer models achieve superior accuracy (up to 98.4%), they impose significant computational overhead compared to lightweight LSTM architectures. The study concludes by proposing a roadmap for future research, emphasizing Explainable AI (XAI) and Federated Learning for privacy-preserving detection.
misinformation are profound, ranging from manipulating political elections to inciting social unrest and spreading public health myths.
Keywords-Fake News Detection, Deep Learning, Natural Language Processing, Transformers, BERT, Bi-LSTM, Multimodal Learning, Generative AI.
II. RELATED WORK
Research by Vosoughi et al. [1] indicates that falsehoods diffuse significantly farther, faster, deeper, and more broadly than the truth in all categories of information. This phenomenon is driven by the "novelty hypothesis," which suggests that fake news is more novel and emotionally evocative than reality. Traditional detection approaches relied on feature engineering, extracting linguistic features such as ngrams, punctuation analysis, and readability scores. However, these methods are brittle and domain-specific. Deep Learning (DL) automates feature extraction, learning hierarchical representations of data. This paper provides a granular analysis of DL techniques, focusing on sequence modelling, attention mechanisms, and the new challenge of AI-generated text.
The field of automated deception detection has evolved through several distinct phases.
NOMENCLATURE
DL: Deep Learning NLP: Natural Language Processing CNN: Convolutional Neural Network LSTM: Long Short-Term Memory Bi-LSTM: Bidirectional LSTM BERT: Bidirectional Encoder Representations from Transformers RoBERTa: Robustly Optimized BERT Pretraining Approach TP/TN: True Positive / True Negative FP/FN: False Positive / False Negative
I. INTRODUCTION The advent of Web 2.0 has democratized content creation, allowing information to bypass traditional editorial gatekeepers. While this fosters free speech, it has concurrently enabled the weaponization of information. "Fake news" is defined as intentionally fabricated information published with the intent to
© 2025, IRJET
|
Impact Factor value: 8.315
|
Phase 1: Statistical Machine Learning (20152018): Early approaches utilized Support Vector Machines (SVM) and Naïve Bayes classifiers. These models relied heavily on "bag-of-words" approaches, which ignored the context and order of words.
Phase 2: Deep Sequence Modeling (20182021): The introduction of Recurrent Neural Networks (RNNs) allowed models to process text as a sequence. Researchers like Ruchansky et al. developed the CSI model, which combined text analysis with user behavior analysis.
Phase 3: The Transformer Era (2021Present): The release of BERT by Google revolutionized NLP. Current state-of-the-art research focuses on fine-tuning pre-trained Transformers to detect subtle nuances, sarcasm, and framing in news articles.
ISO 9001:2008 Certified Journal
|
Page 288