Authors: Research Scholar Priya Mishra, Dr. Deepika Pathak
Abstract: The widespread adoption of generative artificial intelligence has accelerated the creation of highly realistic deepfake images, creating major concerns regarding digital security, misinformation, identity fraud and media authenticity. Traditional detection approaches often struggle to maintain robustness against evolving synthetic image generation techniques and real-world image transformations. This review paper presents a comprehensive analysis of advanced preprocessing strategies and hybrid neural architectures for high-accuracy image deepfake detection. The study critically examines state-of-the-art frameworks including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), attention-based models and hybrid architectures such as GenConViT, and investigates the role of preprocessing techniques including normalisation, facial alignment, edge enhancement, frequency-domain transformation and feature optimisation. Eleven studies are reviewed and organised into three themes covering deep learning architectures, advanced preprocessing and feature enhancement, and challenges with generalisation. The reviewed evidence is consolidated into comparative tables that map each study to its focus, approach, principal finding and limitation, alongside a comparison of architecture families and a summary of preprocessing techniques. Challenges such as adversarial attacks, cross-dataset generalisation, zero-shot detection limitations and computational complexity are discussed, and seven research gaps are identified and mapped to corresponding future research directions. Directions focusing on explainable AI, lightweight hybrid models, generalised feature learning and multimodal detection are presented to support the development of scalable, interpretable and computationally efficient deepfake detection systems.
