Deepfake Image Detection Using ResNet-50: A Systematic Literature Review

Abstract

With the rapid proliferation of deepfake technologies, the generation and propagation of manipulated images pose significant threats to digital security, privacy, and public trust. Convolutional Neural Networks (CNNs), particularly ResNet-50, have gained attention for their robust feature extraction capabilities in deepfake detection tasks. This systematic literature review aims to synthesize recent research efforts that leverage ResNet-50 for the detection of deepfake images. Following PRISMA guidelines, we analyzed peer-reviewed articles published between 2020 and 2025 across major scientific databases including IEEE Xplore, Science Direct, SpringerLink, and ACM Digital Library. Our review explores how ResNet-50 is employed either as a standalone classifier, a transfer learning backbone, or part of hybrid architectures and evaluates its performance across various datasets such as, Flickr / FFHQ FaceForensics++, Celeb-DF, and DFDC. We identify key limitations in current approaches, particularly in terms of generalization to unseen manipulations, dataset biases, and adversarial robustness. The findings indicate that while ResNet-50 demonstrates competitive accuracy and computational efficiency, challenges remain in ensuring model reliability in real-world applications. This review provides a foundational reference for future work aimed at enhancing deepfake image detection systems using deep learning.