Digital images underpin reporting, science, security, and social communication, yet modern editing and generative tools make manipulation effortless and often imperceptible. Digital image forgery detection is fundamentally a pattern recognition problem: the task is to learn and exploit stable patterns in authentic imagery and to recognize distributional departures introduced by manipulations. This review synthesizes research on image forgery detection and localization (IFDL) from 2016 to 2025, spanning traditional passive forensics, CNN-based detectors, and the latest transformer-driven models. We formalize a two-axis taxonomy, using a technical approach (handcrafted / statistical, CNNs, transformers, GAN- and multimodal-assisted) and by forgery type (copy-move, splicing, object removal / inpainting, retouching, and AI-generated fakes). Classical methods exploit lighting, compression, noise, and block / keypoint matching for copy-move and splicing; deep models advance to pixel-wise localization with multi-stream cues (RGB, noise / frequency, edges) and robust pretraining. We analyze publicly used datasets (e.g. Columbia, CASIA, MICC, CoMoFoD, COVERAGE, NIST16, Realistic Tampering, PS-Battles, IMD2020, DEFACTO, tampCOCO, DF2023), highlighting scale, annotation granularity, and manipulation diversity, and relate them to reported performance trends. A critical gap analysis identifies persistent challenges: cross-dataset generalization, robustness to laundering (compression, blur, resizing), explainability and evidentiary utility, data scarcity for subtle edits, and efficiency for large-scale deployment.
Munawar et al. (Wed,) studied this question.