RIFD-NET: A ROBUST IMAGE FORGERY DETECTION NETWORK

Abstract

Image forgery detection has become increasingly important with the widespread use of digital images and editing tools that enable manipulation such as image splicing. Image splicing involves removing or inserting objects into images to create misleading or fake content, making it difficult to identify authenticity through visual inspection. Traditional forgery detection techniques often fail when images contain noise, which significantly reduces detection accuracy. To address this challenge, this project proposes a robust image forgery detection framework based on RIFD-Net (Robust Image Forgery Detection Network). The proposed model integrates multiple stages including noise detection, de-noising, and forgery detection. Initially, a noise detection module identifies different types of noise such as Gaussian and Salt & Pepper noise. An advanced de-noising model (ADNet) is then applied to clean the image and enhance its quality. Following this, a Siamese network is used to compare features between original and test images to detect forgery regions. The model is trained using benchmark datasets such as LIVE1 for noise detection and the Columbia dataset for splice detection. The performance of the system is evaluated using Mean Average Precision (MAP), achieving an accuracy of approximately 89% on test images. The system also localizes forged regions by highlighting them with bounding boxes. This approach significantly improves forgery detection accuracy in noisy environments, making it suitable for applications in digital forensics, media authentication, and cybersecurity.