Semantic-Aware Hierarchical Forgery Network for Image Forgery Detection using Structural and Semantic Inconsistency Learning

Abstract

Image forgery detection is a vital research area because the manipulation of digital images using advanced editing software poses a serious threat to the integrity and credibility of information in various fields, including forensic science, news media, and security. This research proposes an efficient and accurate method for image forgery detection that identifies both structural and semantic inconsistencies introduced during forgery processes, including splicing, copy-move, and image inpainting. The proposed approach presents a Semantic-Aware Hierarchical Forgery Network (SAHF-Net) that combines multi-scale feature learning with dual-branch attention mechanisms to simultaneously analyze local structural inconsistencies and global semantic inconsistencies through multiple convolutional and Transformer layers, enabling effective modeling of inconsistencies across both spatial and semantic contexts. Experimental results on benchmark image forgery datasets demonstrate the effectiveness of the proposed method, achieving 96.3% accuracy in forgery detection. The findings indicate that the model outperforms existing single-stage detection approaches, especially under challenging forgery conditions. In conclusion, the SAHF-Net provides an effective and efficient solution for image forgery detection by jointly capturing structural and semantic inconsistencies, making it highly suitable for forensic applications.