From Detection to Localization: Advances in Image Forgery Analysis with AI

The rise of generative models has intensified the challenge of detecting manipulated images, as these tools create increasingly convincing forgeries, making multimedia forensics a critical field. This thesis provides a complete analysis of state-of-the-art methods for detecting image manipulations. It includes a survey of existing approaches, replication of key experiments to address unclear or underreported findings, and an examination of common limitations and failure modes. Traditionally, image manipulation detection has been treated as a binary classification problem (real vs. fake), which restricts the ability to identify and localize diverse manipulation types. To overcome the limitations of binary classification, this work enhances an already existing detector by developing a unified multiclass framework (real vs. synthetic vs. tampered). This framework not only categorizes the image authenticity but also integrates a segmentation branch for pixel-level localization of tampered regions. The results demonstrate that: (1) replicating prior studies highlights vulnerabilities to preprocessing decisions and dataset inconsistencies, (2) the multiclass approach enables a more granular diagnosis of image authenticity and, when coupled with a dedicated segmentation branch, achieves state-of-the-art localization performance on its primary benchmark, (3) despite these advancements, significant challenges in cross-dataset generalization and robustness to diverse manipulation types persist. The primary contributions of this work include a critical review of the literature, a detailed analysis of failure modes, the development and evaluation of a multiclass detection framework, and a discussion of unresolved challenges and potential future research directions.