Aristotle University Of Thessaloniki (AUTH)
Assessment of information truthlikeliness and authenticity
Abstract
dc:descriptionIn the era of digital information proliferation, the potential disruptive impact on citizens, democracy, and society has raised concerns. To mitigate these risks, it is crucial to assess the authenticity of multimedia content and ensure the dissemination of credible information. This doctoral thesis encompasses three research pillars addressing the challenges of multimedia authentication and information leveraging through novel approaches in spectral analysis, hypergraph theory, and convolutional neural networks. The first part of this thesis focuses on the Electric Network Frequency (ENF) criterion for multimedia authentication. A non-parametric approach for ENF estimation is proposed, incorporating a tailored lag window design into the Blackman-Tukey spectral estimation method. This approach addresses leakage reduction as an energy maximization task within the spectral window's main lobe. Additionally, temporal windowing through the filter-bank Capon spectral estimator is introduced. A Gohberg-Semencul factorization technique is employed for matrix inversions, expediting computations. Challenges of ENF estimation in multimedia recordings, particularly interference from speech content, are addressed. An innovative automated approach for ENF estimation in both static and non-static digital video recordings is proposed, overcoming challenges by exploiting regions with similar characteristics in each video frame. The second research pillar delves into hypergraph learning for personalized image and tag recommendation. To address computational complexities, a novel block randomized Singular Value Decomposition via subspace iteration is integrated into hypergraph learning. An efficient learning scheme for hypergraph ranking is introduced, utilizing multiple optimizations and adaptive hyperedge weight estimation based on gradient descent. The third research pillar differentiates computer-generated images from natural ones through an end-to-end framework using supervised contrastive learning and style transfer via deep neural networks. Hypothesis testing confirms the statistical significance of the improvements achieved. By exploring innovative approaches within these three pillars, this thesis advances multimedia authentication, hypergraph learning, and computer-generated image discrimination, addressing significant challenges in the digital information landscape.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Karantaidis, Georgios
- Καρανταΐδης, Γεώργιος
Subjects
dc:subject × 18- Φασματική ανάλυση
- Επεξεργασία σήματος
- Συνελικτικά νευρωνικά δίκτυα
- Συχνότητα ηλεκτρικού δικτύου
- Εκμάθηση υπεργραφημάτων
- Έλεγχος αυθεντικότητας πολυμέσων
- Multimedia forensics
- Signal processing
- Spectral analysis
- Convolutional neural networks
- Electric network frequency
- Hypergraph learning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57173