Zingarini, Giada (2025) Towards general integrity verification approaches for natural and medical images. [Tesi di dottorato]

Documento PDF zingarini_giada_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (8MB) | Richiedi una copia

Abstract

The rise of generative AI has revolutionized the production of synthetic content, enabling the creation of high-quality and sophisticated media with unprecedented ease, even for users without advanced technical expertise. Through intuitive interfaces and powerful pretrained models, individuals can now generate synthetic images from simple inputs or textual prompts. This accessibility has democratized content creation, however, it also introduces significant challenges, including the potential misuse, such as the large-scale spread of misinformation. These concerns extend beyond natural imagery to encompass medical and scientific images within academic publications. Manipulating such content can alter diagnoses, support unproven theories, or promote misleading assumptions. This thesis provides a comprehensive analysis of this issue across both image domains, proposing general solutions for image forgery detection and localization. Since natural images are often shared repeatedly across social networks, detector robustness becomes a key requirement. The proposed method, B-Free, is specifically designed to address this need, demonstrating strong generalization across a wide range of generative tools. It employs a novel training paradigm that eliminates both coding and semantic content biases from the training data, compelling the detector to focus exclusively on generative artifacts. Similarly, medical and scientific images are also vulnerable to AI-driven manipulation. To address this, a tampered dataset of 3D images, (M3Dsynth), is introduced to enable an extensive benchmark of SoTA detection and localization methods. Furthermore, a Fusion Model is proposed to enhance forgery localization capabilities for scientific imagery. The evaluations highlight the necessity of deploying specific data and models tailored to these different types of manipulations.

Downloads

Actions (login required)

Modifica documento