Advancing Digital Forgery Detection: A Comparative Analysis of U-Net, Buster Net, Efficient B0+ U-Net, and VGG 16+ U-Net Models

Abstract

Forgery Detection in the area of digital images is indispensable for insuring authenticity, particularly in detecting Copy-Move, Splicing. In this research, a meticulous analysis of latest learning architectures i.e. U-Net, Buster Net, Efficient B0+U-Net and VGG16+U-Net are compared for the detecting these types of forgeries. U-Net is popular due to its potent segmentation capabilities to provide pixel-level forgery localization while Efficient B0+U-Net is a lightweight variation, which aims to enhance computational efficiency along with preserving accuracy, enabling faster computation time. VGG16+U-Net utilizes pre-trained VGG 16 model as encoder to extracts hierarchical features and decoder up-samples the features while merging spatial details from the encoder through skip connections to provide faster training, better generalization, better accuracy and faster computation. Buster Net, a leading-edge network particularly designed for forgery detection, employ dual-stream processing to predict manipulated regions in an image effectively. All these models utilize IEEE IFS-TC Image Forensic Challenge dataset and CASIA V2. The evaluation of these models is done based on accuracy, F1, IoU and AUC along with other parameters, computational speed, and robustness to post-processing attacks. Experimental results demonstrate the strengths and limitations of each model, with U-Net and Lightweight Buster Net excelling in complex manipulations in terms of Evaluation metrics, U-Net and Buster Net models have given higher Accuracy (≈97%), F1 score (≈97%) and less computation time for both datasets as compared to other two Models. These findings highlight the trade-offs in model selection for this particular application.