Digital image forgery detection is a vital area of research in digital forensics, aiming to authenticate visual content in the face of increasingly sophisticated manipulation techniques. This paper presents a comprehensive overview of the field, integrating key concepts in its technical landscape. The categorization of detection methods typically includes active approaches that depend on embedded watermarks or signatures, and passive (blind) techniques that analyze the image content itself without prior information. Underlying these are various detection principles, such as identifying inconsistencies in pixel patterns, compression artifacts, illumination, and sensor noise. The paper explores the specific characteristics of detection techniques, analyzing their strengths and limitations. Pixel-based and statistical methods offer efficiency for copy? move and splicing detection but often lack robustness under compression or scaling. Frequency-domain methods and physics-based analysis provide deeper insights, but they can be computationally intensive or sensitive to environmental conditions. The evaluation of detection models is crucial, relying on diverse datasets, realistic manipulation scenarios, and adversarial robustness testing. Effective evaluation metrics include accuracy, precision, recall, F1-score, AUC-ROC and IoU, which collectively assess classification and localization performance. The deep learning approach in forgery detection has significantly advanced the field, with convolutional neural networks and transformer-based models learning complex tampering artifacts. However, challenges persist in forgery detection, including evolving manipulation methods, dataset limitations, explainability concerns, and vulnerability to adversarial attacks. Finally, the authors discuss trends and future directions, such as self-supervised learning, multimodal forensic integration, domain adaptation, and real-time detection frameworks, paving the way for more resilient and scalable forensic tools.
Petar Čisar (Wed,) studied this question.