Ensuring Visual Integrity: Deep Learning-Based Solutions for Authentic Image Forgery Detection

Abstract

Digital image manipulation has become increasingly prevalent with the advancement of image editing tools, posing significant challenges in digital forensics. Detecting and localizing two common types of image forgery copy-move forgery and spliced image forgery remains a critical task. This paper proposes an approach that leverages EfficientFormer for forgery detection and BCU-Net with a spatial attention mechanism for localization. EfficientFormer is used to classify images as forged or original, while BCU-Net precisely identifies and localizes the forged regions. The study utilizes well-known datasets in the field of digital forensics, including CASIA 1.0, CASIA 2.0, and CoMoFoD v2. The results are evaluated using precision, recall and the F1 score for classification with the EfficientFormer model. For localization, training loss, training accuracy, validation loss, validation accuracy, and Intersection over Union (IoU) are used as metrics following the application of BCU-Net. The approach achieves high accuracy rates, including 99.84% for copy-move forgery detection on the CASIA 1.0 dataset, 99.81% for spliced image forgery detection on the same dataset, 99.97% for CASIA 2.0, and 99.96% on the CoMoFoD v2 dataset. A comparison with state-of-the-art methods shows that the proposed approach consistently outperforms existing methods, demonstrating superior performance across all evaluation metrics.