Guillaro, Fabrizio (2024) Towards Robust and General Image Forgery Detection and Localization. [Tesi di dottorato]
Questa è la versione più aggiornata di questo documento.
![[thumbnail of PhDThesisGuillaro_firmato.pdf]](https://www.fedoa.unina.it/style/images/fileicons/other.png "PhDThesisGuillaro_firmato.pdf") |
Documento PDF
PhDThesisGuillaro_firmato.pdf Visibile a [TBR] Amministratori dell'archivio Download (10MB) | Richiedi una copia |
| Tipologia del documento: | Tesi di dottorato |
|---|---|
| Lingua: | English |
| Titolo: | Towards Robust and General Image Forgery Detection and Localization |
| Autori: | Autore Email Guillaro, Fabrizio fabrizio.guillaro@unina.it |
| Data: | 11 Dicembre 2024 |
| Numero di pagine: | 150 |
| Istituzione: | Università degli Studi di Napoli Federico II |
| Dipartimento: | Ingegneria Elettrica e delle Tecnologie dell'Informazione |
| Dottorato: | Information technology and electrical engineering |
| Ciclo di dottorato: | 37 |
| Coordinatore del Corso di dottorato: | nome email Russo, Stefano stefano.russo@unina.it |
| Tutor: | nome email Verdoliva, Luisa [non definito] Poggi, Giovanni [non definito] |
| Data: | 11 Dicembre 2024 |
| Numero di pagine: | 150 |
| Parole chiave: | Image Forensics, Forgery Detection, Forgery Localization, Deepfakes, Synthetic Image Detection, Adversarial Robustness |
| Settori scientifico-disciplinari del MIUR: | Area 09 - Ingegneria industriale e dell'informazione > ING-INF/03 - Telecomunicazioni |
| Informazioni aggiuntive: | ITEE 37th cycle |
| Depositato il: | 29 Dic 2024 09:46 |
| Ultima modifica: | 12 Ago 2026 05:37 |
| URI: | https://www.fedoa.unina.it/id/eprint/16429 |
Available Versions of this Item
-
Interazioni tra olio extravergine di oliva e matrici alimentari nei processi di conservazione e cottura. (deposited 24 Giu 2008)
- Towards Robust and General Image Forgery Detection and Localization. (deposited 29 Dic 2024 09:46) [Attualmente visualizzato]
Abstract
Synthetic media generation has seen tremendous progress in the span of just a few years. Powered by large language models, text-to-image synthesis tools allow the user to create and modify images at will by means of simple text instructions. Although this represents a great opportunity for visual arts applications, it can also be used by malicious actors who aim at deliberately disseminating disinformation. Consequently, multimedia forensics has been drawing increasing attention, with a strong demand for effective detectors able to distinguish manipulated images from real ones. Aim of this thesis is to develop a robust and general approach for detecting and localizing image manipulations. The method is based on the extraction of a camera model fingerprint (Noiseprint++) designed to improve robustness to post-processing operations that may attenuate forensic traces. To this end, a contrastive learning approach is applied on image patches that underwent a different combination of processing operations (editing history), with a training set consisting only of real images. Hence, forgeries are detected as deviations from the expected regular pattern that characterizes each pristine image. TruFor, the proposed framework, combines the Noiseprint++ with the RGB image to provide as output a confidence map that highlights areas where localization predictions may be error prone. This is particularly important in forensic applications to reduce false alarms. Additionally, in this thesis, the robustness of current forensic detectors to adversarial attacks has been explored. The transferability of attacks between different families (CNNs, ViTs) has been analyzed, both numerically and with the help of the resulting Fourier-domain patterns. This analysis sheds light on how forensic detectors work and is therefore a valuable tool for developing more effective and robust methods.
Downloads
Downloads per month over past year
Actions (login required)
 |
Modifica documento |
