Deep Learning Techniques for Image Plagiarism Detection: A Systematic Review

Image plagiarism is unauthorized copying, editing or reusing of digital images without due permission or citing. As social media, online publishing, and Artificial Intelligence (AI)-based tools to edit and remove plagiarism in images have developed at an alarming pace, it has become more difficult to detect image plagiarism. It provokes some serious questions, associated with the copyright violations, intellectual property and the misuse of visual materials ethically. The objective of this review is to explore the increasing challenge of image plagiarism in the time of artificial intelligence and assess deep learning techniques implemented for unauthorized image reuse detection. A systematic, structured Preferred Reporting Items for Systematic reviews and Meta-Analyses aligned methodology was used to collect, screen and analyze the studies published between 2015 and 2025 including traditional models, machine learning models and deep learning models. The feature-based techniques, Convolutional Neural Networks (CNN) architectures, Siamese Networks, Generative Adversarial Networks based detectors, and hybrid multimodal systems are discussed in the review. Key findings highlight the results show that the performance of deep CNN, especially Visual Geometry Group, Residual Network, and Vision Transformers, is largely accurate (94-99%) and also shows good robustness to attack such as cropping, scaling, and color manipulation. However, the analysis shows some significant gaps such as the lack of standardized datasets and the lack of explainability in AI-based decisions. The article concludes that deep CNNs are the most reliable models to be used in image plagiarism detection. Unlike prior surveys limited to architectural overviews, this work uniquely integrates empirical dataset benchmarking across 50+ sources and proposes interpretable multimodal frameworks as the path forward for reliable plagiarism detection. The novelty of this review is to present a complete taxonomy, research gaps, and the necessity of having interpretable and multimodal standardized detection frameworks.