Research on the identification method of JPEG image copy-paste forgery based on BAG optimization and lightweight GNN

Abstract

A forgery detection method based on block artifact grid (BAG) optimization and lightweight graph neural network (GNN) is proposed to address the problems in JPEG image copy-paste forgery detection, such as insufficient robustness to compression interference in existing methods, difficulty in balancing lightweight and detection accuracy in deep learning models, and inadequate feature mining between image blocks. This method first extracts the 8×8 DCT coefficient statistical features of JPEG images and completes normalization, then maps each 8×8 DCT block feature to an independent graph node, constructs edges based on spatial adjacency and feature similarity, and adopts a spatial neighborhood-constrained graph topology; then constructs the BAG structure and strengthens the feature association of the forged region through attention weight optimization and redundant node clipping, with clear attention function form, optimization objective and pruning criterion disclosed; finally uses lightweight GNN a custom lightweight GCN variant to achieve efficient feature learning and forged region discrimination. The proposed method fully combines the advantages of JPEG compression characteristics and graph structure association learning, taking into account detection accuracy, inference efficiency and anti-compression robustness, and is verified under JPEG quality factors Q=50-95 with single compression; can effectively complete the precise identification and region localization of JPEG image copy-paste forgery.