Abstract
<title>Abstract</title> <p>Image manipulation localization (IML), the task of accurately identifying the manipulated regions within an image, remains a significant challenge. Conven tional deep learning methods often struggle with generalization and adapting to new manipulation schemes. To address these limitations, a novel two-stage IML system is proposed. The first stage employs a Dual-Stream Manipulation Classifier, fusing features from the standard RGB domain with low-level foren sic noise artifacts extracted via Steganalysis Rich Model (SRM) filters using a ResNet like 4-stage backbone, to accurately classify the input image into one of four common manipulation types: Copy-Move, Splicing, Removal (Inpainting), or Enhancement. In the second stage, the classified image is directed to a spe cialized Conditional Diffusion Model (CDM), which treats the localization task as an image-to-mask generation problem. These CDMs are conditioned on the input image, and are trained with a modified loss function incorporating both Mean Squared Error (MSE) and Intersection over Union (IoU) to better handle smaller manipulation masks and encourage better spatial structure. Evaluated on the DF2023 dataset, the manipulation classifier achieves an accuracy of 89% and localization CDMs achieve average IoU of 0.70 and F1-Score of 0.77. The system demonstrates competitive performance on benchmark datasets, confirm ing the efficacy of a modular, generative approach for tackling the complexity of image forgeries.</p>