A Systematic Analytical Review of Machine Learning Models for Forgery Detection and Localization in Image Document

The rapid development of image editing tools and the rise in digital images forgery creates obvious threats, which is severe regarding the preservation of visual data integrity. Hence prompt detection and localization techniques are needed in coming time. Although numerous methodologies have been proposed, past reviews suffer from partial fragmentation of analysis, their explorations of cross-dataset generalization can be sparse and there is limited emphasis on hybridizing disentangled features and lightweight models resulting in the absence of a comprehensive taxonomy for image forgery detection approaches. In order to tackle these shortcomings, this paper presents a comprehensive overview and comparative analysis of stateof- the-art forgery detection and localization systems, by putting forward a structured taxonomy according to their architectures, applications and limitations. The review covers multiple methodologies, including transfer learning-based approaches (MobileNetV2), hybrid deep learning frameworks (FBI-Net and DIF-Net), and transformer-enhanced architectures (TransU2-Net), with assessments of accuracy, F1 score, robustness, and scalability.In addition to that, it outlines models working on domain generalization challenges, such as the CFM Framework, AUC > 0.96, along with adversarial robustness issues such as Inspector and MDCF-Net. It will helps the field by bringing everything together under one system to check how well models work, find the best ways to detect certain types of forgery, and create new hybrid and explainable AI models.