Deefake Detection, Attribution, and Authentication: Insights from the FF4ALL Project

The rapid advancement of deep generative models has enabled the large-scale creation of highly realistic deepfakes. While these technologies support innovative applications, they also pose serious threats to trust, security, and digital integrity. As a response, the FF4ALL project investigates deepfake media forensics through a unified framework that integrates source attribution, passive detection, robustness analysis in realistic conditions, and active authentication mechanisms. This paper provides a consolidated overview of the main scientific outcomes achieved within FF4ALL. On the attribution side, novel hierarchical and open-world strategies are presented to identify both the generation technology and the specific model instance responsible for synthetic content. For passive detection, the project advances state-of-the-art methodologies in audio, visual, and multimodal domains, with particular emphasis on generalization to unseen attacks, adversarial robustness, and explainability. Realistic deployment scenarios are addressed through extensive evaluation under social-media compression, continual learning, and out-of-distribution conditions. Beyond passive analysis, FF4ALL develops active authentication solutions, including geometry-aware forensic features, fragile watermarking, cryptographic croppable signatures, and blockchain-based timestamping.

Recommended citation: Amerini, I. et al. "Deefake Detection, Attribution, and Authentication: Insights from the FF4ALL Project.", Proceedings of the Joint National Conference on Cybersecurity (ITASEC & SERICS 2026)
Download Paper | Download Bibtex