Video-based Morphing Attack Detection

Face morphing represents a real threat to facial recognition systems, as it enables the generation of hybrid images of two subjects capable of evading biometric controls, both human and automatic, adopted in critical contexts such as airport gates. In this thesis first the main morphing attack detection (MAD) generation and detection techniques, both S-MAD (single-image) and D-MAD (differential), are analyzed in detail. A new and novel paradigm, called V-MAD (Video-based Morphing Attack Detection), which exploits video sequences instead of single static images to improve the effectiveness of detection, is then introduced and explored. After analyzing various MAD score fusion strategies and integration with image quality metrics, it is described in detail how it was possible to build a proprietary dataset conforming to ICAO standards, characterized by excellent quality and good variability and heterogeneity of the data present, and experiments were conducted on it and on public and semi-public databases. The results obtained show that V-MAD, even when implemented with simple aggregation techniques, has the potential to be able to outperform traditional D-MAD systems, with further improvement achieved through the use of Face Image Quality Assessment algorithms. This work therefore sets the stage for a new direction in biometric security research, opening concrete prospects for the adoption of more robust control systems that are resilient to morphing attacks.