An Attention U-Net Based on Noise Inconsistency for Robust Image Forgery Detection and Precise Pixel-Level Localization

With the rapid advancement of image editing tools, digitally manipulated images have become increasingly common, raising serious concerns about the authenticity and reliability of visual information across domains such as social media, journalism, healthcare, and legal investigations. Although recent deep learning approaches, especially U-Net and attention-based models, have demonstrated promising results in forgery detection and localization, many of these methods mainly depend on visual and texture-based features. As a result, they often struggle to maintain performance when images undergo post-processing operations like compression, resizing, or contrast adjustments. To address these challenges, this work introduces a Noise-Inconsistency Guided Attention U-Net (NIGA-UNet) for robust image forgery detection and precise pixel-level localization. The proposed model focuses on capturing noise-related inconsistencies by extracting noise residuals and analyzing local variations, which are then incorporated into a noise-guided attention mechanism. This enables the network to better highlight manipulated regions. In addition, a dual-encoder U-Net architecture with multi-scale feature fusion is employed to effectively combine both visual and noise-domain information.