An Attention-Enhanced Multi-Scale Framework for Copy–Move Forgery Detection and Localization

Received: 25 February 2026 | Revised: 16 March 2026 | Accepted: 24 March 2026 | Online: 5 April 2026

Corresponding author: Vanishri. V. Sataraddi

Abstract

When a region in an image is copied and pasted elsewhere within the same image, the resulting forgery is considered a Copy–Move Forgery (CMF). CMF is a challenging problem in digital image forensics, as the copied region retains the original image's statistical properties and can also be manipulated by rotation, scaling, compression, and other transformations. In this paper, we propose an attention-guided deep learning framework for robust Copy–Move Forgery Detection (CMFD) with precise localization. The proposed network architecture aggregates hierarchical convolutional feature maps and boosts the representational saliency through successive channel and spatial attention modules. A multi-scale decoder with integrated self-correlation computation and a lightweight transformer block further encourage the extraction of long-range duplicated patterns, enhancing localization accuracy. Extensive experiments on benchmark datasets validate that this method achieves up to 95% precision, 93% recall, and 94% F1-score, significantly outperforming notable baseline methods. The network also demonstrates robust generalizability against commonly encountered postprocessing operations such as JPEG compression, noise, and moderate geometric transformations. The complete architectural design and mathematical formulation are provided to ensure reproducibility and practical applicability in real-world image forensic scenarios.