Two major European research projects focused on deepfake detection and mitigation—FF4ALL and FUN-Media—have ended. These initiatives represent strategic efforts to address the growing impact of manipulated digital content—fake images, videos, and audio—on information security.
The Image and Sound Processing Lab (ISPL) of the Department of Electronics, Information and Bioengineering at Politecnico di Milano, also supported by Italy’s PNRR funding, played a key role in both projects. Within FF4ALL, researchers focused on analysing emerging phenomena related to the generation of synthetic images and videos, while in FUN-Media they addressed the increasingly critical issue of audio deepfakes.
The results mark a significant step forward in the development of reliable technologies to safeguard digital information, counter disinformation, and protect users in an increasingly complex and dynamic media ecosystem.
In the FF4ALL project, the ISPL team, led by Prof. Stefano Tubaro, investigated how fake images and videos are engineered and disseminated. In particular, they analyzed advanced techniques capable of transforming real content into highly realistic synthetic versions, making authenticity verification more challenging and concealing key traces needed for forensic analysis. In parallel, the researchers developed new tools for detecting synthetic faces by combining three-dimensional geometric information with structural facial features. These solutions improve the ability of models to identify fakes, maintaining strong performance even after post-processing operations such as compression or editing.
In the FUN-Media project, coordinated by Prof. Paolo Bestagini, the focus shifted to audio deepfake detection—one of the most insidious emerging threats in the digital security landscape. To address this challenge, the team developed new architectures based on “Mixture of Experts” models, capable of combining multiple specialized systems to enhance performance even when facing generative techniques not encountered during training. These approaches offer greater flexibility and adaptability than traditional detectors, proving particularly effective in complex and rapidly evolving scenarios.
An additional line of research explored the use of forensic detectors based on anomaly detection. In this approach, models are trained exclusively on authentic speech signals, learning their distinctive characteristics and identifying synthetic content as deviations from expected behaviour.