Image manipulation detection (IMD) aims to determine whether an image has been tampered with and to identify the manipulated regions. These capabilities have become increasingly important with the rapid advancement of media editing and generation technologies, such as Photoshop and generative AI methods, which underscore the need for robust tools for media authentication. Although current state-of-the-art (SoTA) methods achieve strong results on common manipulation types, such as splicing, copy-move, and removal, they often struggle to generalize to manipulation types not represented in the training data. Consequently, their real-world applicability remains limited, with performance degrading significantly in practical scenarios. In this dissertation, we advance image manipulation detection through three key contributions. First, we introduce the Challenging Image Manipulation Detection (CIMD) benchmark, a novel, high-quality dataset with fine-grained annotations designed to evaluate SoTA methods on both editing-based and compression-based manipulations under realistic and more complex conditions. Using CIMD, we demonstrate that existing SoTA methods struggle with small tampered regions and double-compression cases with identical quality factors, and we propose a two-branch HRNet-based model that significantly outperforms prior work on these challenges. Second, we present a unified unsupervised and weakly supervised framework that reduces reliance on pixel-level annotations. This framework leverages implicit neural representations and selective contrastive learning, achieving detection performance comparable to supervised methods while improving robustness to unseen manipulations. Finally, we develop a training-free diffusion-based approach that exploits inconsistencies between conditional and unconditional reconstructions for manipulation detection. This method requires no external training data and outperforms existing unsupervised and weakly supervised techniques, while achieving competitive results with fully supervised models across multiple benchmark datasets. Collectively, these contributions strengthen IMD performance in realistic tampering scenarios and broaden its applicability to forensic settings where manipulation types are diverse, rapidly evolving, and often unseen during training.
Z. H. Zhang (Wed,) studied this question.