A Comprehensive Review of Machine Learning and Deep Learning Approaches for Fake News Detection

The exponential rise of online content has fueled the widespread circulation of fake news, posing significant threats to public trust, societal stability, and information authenticity. Detecting fake news has therefore become a critical research domain across natural language processing (NLP), computer vision, multimodal learning, and data mining. This review paper presents a comprehensive survey of machine learning (ML), deep learning (DL), and transformer-based approaches designed for identifying misleading or deceptive content across digital platforms. Traditional ML techniques were initially deployed using handcrafted linguistic and statistical features; however, their limited contextual understanding restricted their effectiveness. Deep neural networks—including CNNs, RNNs, LSTMs, and hybrid architectures—marked a shift toward automated feature extraction, enabling models to capture syntax, semantics, sentiment, and long-range dependencies. More recent advancements leverage transformer-based architectures (BERT, RoBERTa, XLNet), multimodal fusion models, and cross-modal attention mechanisms that analyze both textual and visual cues to detect sophisticated forms of misinformation. This survey evaluates the strengths, limitations, and performance trends across these paradigms, highlights key benchmark datasets, and identifies persistent challenges such as multimodal inconsistencies, evolving linguistic structures, code-mixing, data scarcity, and adversarial misinformation strategies. The paper concludes by proposing future research directions including contrastive learning, GPU-efficient multimodal models, fact-checking augmentation, and explainable AI frameworks for building robust, transparent, and scalable fake news detection systems.