Hybrid Multi-Scale Deep Feature Fusion Network for Retouch Image Forgery Detection

Abstract

The number of digitally altered images is growing because of increasingly powerful and easy-to-use tools for image editing. This development has rendered web-based visual information less credible and more challenging to substantiate. Among various image editing techniques, image retouching is one of the most commonly used methods. Unlike the normal image editing, image retouching does not modify the underlying structure of the image. This procedure can help to alter the texture, tone, and blemishes on the skin without changing the structure of the image. These subtle changes can pose difficulties for traditional forensic methods to detect image manipulation that involves retouching. To overcome these shortcomings, this paper introduces a hybrid deep feature fusion (HDFF) network to detect digital image manipulation (DIM) with retouching forgery. The proposed architecture is a convolutional layered approach with multi-scale technique, which is able to effectively capture subtle texture variations to learn them. The model is assessed against the CASIA v2.0 and FaceForensics++ benchmark datasets, with the standard metrics of accuracy, precision, recall, and F1-score. The results show that the approach attains a detection accuracy of 97.2%, performing better than several existing deep learning-based forgery detection methods. This improvement shows the value of combining multi-level features for the detection of minor retouching artifacts. Also, the approach provides a reliable platform for multimedia authentication and social media validation in digital image forensics.