Final Examination of the Industrial PhD Programme in Big Data and Artificial Intelligence Ordinary Session 21 April 2026 PhD Vittorio Stile
Description
This document contains the doctoral research materials and presentation corresponding to the final PhD thesis defense of Vittorio Stile, titled AI-generated Deepfakes: Detection and Bias Analysis. The research was conducted within the Industrial PhD Programme in Big Data and Artificial Intelligence, Curriculum in Big data management for the digital transition, Cycle XXXVIII, at Universitas Mercatorum, Università telematica delle Camere di Commercio Italiane. The project was supervised by Tutor Prof. Roberto Caldelli and Co-tutor Dott.ssa Elena Santi, funded under the Italian PNRR Scholarship scheme D.M. 352/2022 and co-founded by PricewaterhouseCoopers Business Services SRL.
The three-year academic journey followed a highly collaborative and multidisciplinary approach. During the first two years, the activity integrated university scientific research with industrial application at PwC Italia Srl. Subsequently, between the second and third years, the project saw a significant international expansion through a research period abroad at the University of Cádiz in Spain. This experience allowed the initial objectives, focused on standard neural network-based deepfake detection methods, to evolve into an advanced framework centered on the study of visual attributes, dataset vulnerabilities, and model interpretability.
The core theme and main scientific contribution of the thesis concern the systematic investigation of how high-level facial attributes can influence the performance and fairness of DeepFake detection models. As multimedia forgery generators achieve a level of realism indistinguishable to human perception, automated detectors frequently exhibit hidden vulnerabilities and biases inherited from training data. The study analyzes in detail the direct correlation between the misclassifications of a pre-trained frame-level detector and the visual characteristics of the subjects. Each video in the FaceForensics++ dataset was automatically annotated with additional labels relating to gender, hair color, hair length, ear visibility, and ethnicity, leveraging a semi-supervised facial-attribute recognition pipeline.
To isolate generalization bias, the research included training sessions with controlled exclusions while maintaining a unified test set. The experimental results demonstrate that, compared to a baseline configuration without exclusions which achieves an accuracy of 0.806 and an AUC of 0.823, excluding samples with visible ears causes the most pronounced performance degradation, dropping accuracy to 0.741 and AUC to 0.763. Conversely, excluding subjects with non-visible ears shows a much milder effect, with an accuracy of 0.813 and an AUC of 0.832. Hair length highlights a moderate but consistent impact that interacts directly with ear visibility. The study also explains the asymmetry found in the confusion matrix as a consequence of fixed score thresholds and video-level k-of-n aggregation rules. The emerging evidence demonstrates that ear visibility constitutes a critical factor for robust discrimination between real and fake, motivating the adoption of attribute-aware training strategies such as targeted data curation, attribute-specific feature augmentation, and threshold calibration. The proposed framework offers operational guidelines to mitigate bias and supports the development of more interpretable and operationally reliable DeepFake detection systems.
Other (Italian)
Questa pagina raccoglie i materiali di ricerca e la presentazione relativi alla difesa finale della tesi di dottorato di Vittorio Stile, intitolata AI-generated Deepfakes: Detection and Bias Analysis. Il lavoro è stato sviluppato nell'ambito del Corso di Dottorato Industriale in Big Data e Artificial Intelligence, Curriculum in Big data management for the digital transition, XXXVIII Ciclo, presso l'Universitas Mercatorum, Università telematica delle Camere di Commercio Italiane. La ricerca è stata supervisionata dal Tutor Prof. Roberto Caldelli e dalla Co-tutrice Dott.ssa Elena Santi, finanziata tramite borsa di studio PNRR D.M. 352/2022 e co-finanziata da PricewaterhouseCoopers Business Services SRL.
Il percorso accademico triennale ha seguito un approccio fortemente multidisciplinare e collaborativo. Durante i primi due anni, l'attività ha integrato la ricerca scientifica universitaria con l'applicazione industriale presso PwC Italia Srl. Successivamente, tra il secondo e il terzo anno, il progetto ha visto un'importante estensione internazionale attraverso un periodo di ricerca all'estero presso l'Università di Cadice in Spagna. Questa esperienza ha permesso di fare evolvere gli obiettivi iniziali, focalizzati sui metodi standard di rilevamento dei deepfake basati su reti neurali, verso un framework avanzato incentrato sullo studio degli attributi visivi, sulle vulnerabilità dei dataset e sull'interpretabilità dei modelli.
Il tema centrale e il principale contributo scientifico della tesi riguardano l'indagine sistematica su come gli attributi facciali ad alto livello possano influenzare le prestazioni e l'equità dei modelli di DeepFake detection. Poiché i generatori di falsi multimediali raggiungono un realismo indistinguibile per la percezione umana, i rilevatori automatici mostrano spesso vulnerabilità e bias nascosti ereditati dai dati di addestramento. Lo studio analizza in modo dettagliato la correlazione diretta tra gli errori di classificazione di un rilevatore a livello di frame pre-addestrato e le caratteristiche visive dei soggetti. Ciascun video del dataset FaceForensics++ è stato annotato automaticamente con etichette aggiuntive relative a genere, colore dei capelli, lunghezza dei capelli, visibilità delle orecchie ed etnia, sfruttando una pipeline di riconoscimento semi-superata degli attributi facciali.
Per isolare il bias di generalizzazione, la ricerca ha previsto sessioni di addestramento con esclusioni controllate mantenendo un set di test unificato. I risultati sperimentali dimostrano che, rispetto a una configurazione base senza esclusioni che raggiunge un'accuratezza di 0.806 e un AUC di 0.823, l'esclusione dei campioni con orecchie visibili provoca il degrado prestazionale più marcato, facendo scendere l'accuratezza a 0.741 e l'AUC a 0.763. Al contrario, l'esclusione di soggetti con orecchie non visibili mostra un effetto molto più lieve, con un'accuratezza di 0.813 e un AUC di 0.832. La lunghezza dei capelli evidenzia un impatto moderato ma costante che interagisce direttamente con la visibilità delle orecchie. Lo studio spiega inoltre l'asimmetria riscontrata nella matrice di confusione come conseguenza delle soglie di punteggio fisse e dell'aggregazione a livello video di tipo k-of-n. Le evidenze emerse dimostrano che la visibilità delle orecchie costituisce un fattore critico per una discriminazione robusta tra reale e falso, motivando l'adozione di strategie di addestramento attente agli attributi, come la cura mirata dei dati, l'aumento specifico delle caratteristiche facciali e la calibrazione delle soglie. Il framework proposto offre linee guida operative per mitigare i bias e supporta lo sviluppo di sistemi di DeepFake detection più interpretabili e affidabili sul piano operativo.
Files
2026-04-21_PhD_thesis_defense_STILE_Vittorio.pdf
Files
(18.1 MB)
| Name | Size | Download all |
|---|---|---|
|
md5:0623f90a8309f0e35f459044ba31930d
|
18.1 MB | Preview Download |
Additional details
Related works
- Is described by
- Publication: 10.3389/fimag.2026.1846377 (DOI)
Funding
- European Union
- European Union – NextGeneration EU National Recovery and Resilience Plan (PNRR), Mission 4, Component 2 "From Research to Business"– Investment 3.3 "Introduction of innovative PhD programmes that meet the innovation needs of enterprises and promote the recruitment of researchers by companies" CUP D83C22001880003
Dates
- Issued
-
2026-04-21This is the presented version of my PhD's thesis defense speach
Software
- Repository URL
- https://github.com/vstile/deepfake-attribute-detection
- Programming language
- Python
- Development Status
- Active
References
- Stile, V., Caldelli, R., Balderas-Díaz, S., Guerrero-Contreras, G., & Medina-Bulo, I. (2026). Facial Attribute-Aware DeepFake Detection through Semi-Supervised Facial Attribute Labeling. Frontiers in Imaging, 5. https://doi.org/10.3389/fimag.2026.1846377
- Stile, V., Caldelli, R., Guerrero-Contreras, G., Balderas-Díaz, S., & Medina-Bulo, I. (2025). Analysis of DeepFake Detection through Semi-Supervised Facial Attribute Labeling. Proceedings of the 11th Spanish-German Symposium on Applied Computer Science (SGSOACS 2025), Communications in Computer and Information Science (CCIS), Applied Computer Science(1), 73–88. https://doi.org/10.1007/978-3-032-14816-2
- Benelli, F., Maciariello, F., & Salvadori, C. (2024). The influence of technologies on organizational culture in innovative SMEs. Journal of Robotics and Automation Research, 5(3), 01–11. https://doi.org/10.33140/JRAR
- Stile, V., & Fontanella, A. (2026). AI-Enhanced Building Information Modeling and Big Data Analytics for Civil Engineering Innovation. Journal of Material Sciences & Applied Engineering, 5(2), 01–07. 10.63620/MKJMSAE.2026.1067 https://mkscienceset.com/articles_file/283-_article1777717525.pdf
- Stile, V. (2020). User Experience for Mobile Devices [Università Guglielmo Marconi]. https://doi.org/10.5281/ZENODO.17046468
- Stile, V. (2026). AI-generated Deepfakes: Detection and Bias Analysis [Universitas Mercatorum]. https://doi.org/10.5281/zenodo.19926135
- Stile, V. (2022). Blockchain: Analysis and Realisation of a Transaction. https://doi.org/10.5281/ZENODO.17058889
- Stile, V., Bonino, V., & Cosmo, N. (2024). The impact BI and AI on traditional structures with legal and philosophical insights. Proceedings of the 21st Conference of the Italian Chapter of AIS (itAIS 2024), 21. https://doi.org/10.979.1282308/007
- Stile, V., & Fontanella, A. (2025a). AI-Enhanced Building Information Modelling and Big Data Analytics for Civil Engineering Innovation. Book of Abstract of the 4th International Conference Creativity And Innovation In Digital Economy, Section 1: Innovative open business models and platforms.
- Stile, V., & Fontanella, A. (2025b, May 12). Integrating AI and BIM: Innovations in Civil Engineering through Smart Design and Real-Time Analytics. The 2nd World Conference on Construction and Building Technology (BuildTech Week 2025). https://doi.org/10.5281/ZENODO.17287941
- Sahli, S., Stile, V., & Gillet, D. (2026). Rethinking Higher Computer Science Education in the Age of AI: Insights from Computer Science Students in Tunisia. Proceedings of the 17th IEEE Global Engineering Education Conference (EDUCON 2026), 01–04. https://doi.org/10.1109/EDUCON67543.2026.11574149
- Maciariello, F., Benelli, F., Sangiuolo, G., Lorenzi, E., Caponio, C., & Salvadori, C. (2025). TrackOne: Smart Logistics for a Sustainable and Interoperable Agricultural Supply Chain in the Era of Digitization. 2025 International Conference on Software, Telecommunications and Computer Networks (SoftCOM), 1–7.
- Maciariello, F., Avolio, F., Cicoira, V., Cosmo, N., Laudonia, A., Giannetti, I., Liberanome, P., & Stile, V. (2025, October 29). Competences for Society 5.0: Multidisciplinary Corporate Training for Inclusion, Safety and Competitiveness. Proceedings of the 35th International Conference on Rethinking Services for Society 5.0: Opportunities and Challenges (RESER 2025). The 35th International Conference on Rethinking Services for Society 5.0: Opportunities and Challenges. https://reser.net/2025-roma-35rd-reser-international-conference/
- Liberti, F., Avolio, F., Cicoira, V., Cosmo, N., Laudonia, A., Maciariello, F., & Stile, V. (2025, October 17). Distributed Artificial Intelligence and Health Governance: A Multidimensional Analysis of the Tensions Between Rules, Ethics and Innovation. Proceedings of the 22nd Conference of the Italian Chapter of the Association for Information Systems (ITAIS 2025). The 22st Conference of the Italian Chapter of AIS. https://aisnet.org
- Laudonia, A., Avolio, F., Cosmo, N., Giannetti, I., Liberanome, P., Maciariello, F., & Stile, V. (2025, November 13). AI-Driven Financial Risk Prevention: The Role of HR Analytics in Corporate Crisis Management Under Industry 5.0. Proceedings of the 7th International Conference on Industry of the Future and Smart Manufacturing (ISM 2025). The 7th International Conference on Industry of the Future and Smart Manufacturing (former International Conference on Industry 4.0 and Smart Manufacturing). http://www.sciencedirect.com/
- Fanale, R., Liberti, F., & Stile, V. (2025). Explainable Federated Learning for Secure Telemedicine: Protecting Patient Identity through Privacy-Preserving Deepfake Detection in Digital Health Platforms. Book of Abstract of the 4th International Conference Creativity And Innovation In Digital Economy, Section 3: Challenges of Artificial Intelligence in business.
- Benelli, F., Maciariello, F., & Stile, V. (2026, February 11). Secure Cognitive Orchestration Framework for Multi-Domain Physical Internet: Integrating AI-Driven Logistics, Energy Distribution, and Cybersecurity. Proceedings of the 9th International Conference on Human Intelligent Systems Integration (IHSI 2026). 9th International Conference on Human Intelligent Systems Integration (IHSI 2026): Disruptive and Innovative Technologies, Università di Firenze, Florence, Italy.
- Benelli, F., Maciariello, F., Salvadori, C., Kelliçi, E., & Stile, V. (2025, October 18). Human-AI Collaboration in SMEs: A Role-Sensitive Framework for Cognitive Enterprise Hubs. Proceedings of the 22nd Conference of the Italian Chapter of the Association for Information Systems (ITAIS 2025), Springer LNISO Series. The 22st Conference of the Italian Chapter of AIS. https://aisnet.org
- Benelli, F., Maciariello, F., Marku, R., & Stile, V. (2025). Towards an Energy Physical Internet: Open Business Models and Platforms for Electricity Distribution Enabled by IoT, Blockchain, and Conditional Payments. Book of Abstract of the 4th International Conference Creativity And Innovation In Digital Economy, Section 4: New Pathways in Knowledge, Education and Law.
- Benelli, F., Kellaçi, E., Maciariello, F., & Stile, V. (2025). Artificial Intelligence for Decentralized Orchestration in the Physical Internet: Opportunities, Business Trade-offs, and Risks in Road Freight Logistics. Book of Abstract of the 4th International Conference Creativity And Innovation In Digital Economy, Section 2: Co-creation, living labs and innovation ecosystems.
- Benelli, F., Kellaçi, E., Maciariello, F., Salvadori, C., & Stile, V. (2025). Enhance Student Well-being and Digital Literacy with Machine Learning and Spatial Analysis. Proceedings of the 2nd International Workshop on Education for Artificial Intelligence (EDU4AI 2025), AI*IA SERIES, 4114, Session S2: AI Literacy and Education. https://ceur-ws.org/Vol-4114/
- Benelli, F., Giannetti, I., Stile, V., & Maciariello, F. (2026, January 22). AI-Enabled People & Culture: Un framework socio-tecnico per la sostenibilità organizzativa. Proceedings del 41° Convegno Nazionale dell'Accademia Italiana di Economia Aziendale (AIDEA 2026). 41° Convegno Nazionale dell'Accademia Italiana di Economia Aziendale (AIDEA 2026): Le Intelligenze Aziendali per la competitività sostenibile e il bene comune, Università Cattolica del Sacro Cuore.