The proliferation of generative AI techniques for image generation and editing presents a multitude of challenges, including the dissemination of fraudulent content, the propagation of disinformation, and the erosion of public trust in digital media. While before the onset of this thesis, AI-powered systems demonstrated effectiveness in authenticating images within controlled environments. They exhibited limited capability in addressing the complexities of real world applications.
This thesis responds to these challenges by contributing to the development of forensic systems capable of operating effectively in uncontrolled environments, commonly referred to as ”in the wild”. The initial focus is on tackling the dataset mismatch problem, where in test samples undergo post-processing pipelines or generated by new generative AI tools distinct from those encountered during system training. We introduce a Siamese network for detecting AI-synthetic images and a hybrid architecture, enhancing generalization and robustness against image processing operations. In the second part, we focus on the open-set scenario, devising solutions for synthetic image attribution and facial attribute classification. We develop classifiers with a rejection option, employing hybrid architectures and novel frameworks alongside a verification approach leveraging contrastive learning. These contributions fortify image authentication in uncontrolled environments, mitigating risks of fraud and disinformation.
