Abstract
<title>Abstract</title> <p>Purpose: The widespread availability of advanced photo editing tools has cre ated a critical need for robust digital image forensics. Active Image Manipulation Detection and Localization (AIMDL) offers highly accurate detection and pre cise spatial localization compared to passive forensic methods. However, AIMDL faces an inherent trade-off between watermark transparency (invisibility) and robustness against routine transformations. Methods: This paper proposes a novel AIMDL framework that utilizes TransUNet architectures for the genera tion and recovery of watermarks, coupled with a multi-head UNet-based Forensic Module to handle image authentication, tampering detection, and spatial local ization. The training pipeline integrates Benign Noise and Manipulation modules alongside a carefully curated set of loss functions to optimize watermark quality and forensic accuracy. Models were trained at different watermark scaling fac tors (0.10 and 0.05), the framework allows users to dynamically balance forensic performance and perceptual quality at test time. Results: Experimental results confirm that the proposed pipeline achieves high authentication and detection accuracy (> 0.95) and precise pixel-level localization (IoU > 0.85) of malicious edits while remaining robust against benign transformations like compression and scaling. Conclusion: The framework offers reliable active image forensics, allowing users to tune the balance between imperceptibility and forensic strength according to application-specific security demands.</p>