Core Concepts
Print-and-scan processing can significantly increase the effectiveness of face morphing attacks against face recognition systems, regardless of whether the input image is a morphed or bona fide face.
Abstract
The paper investigates the impact of print-and-scan processing on the effectiveness of face morphing attacks in heterogeneous evaluation scenarios. Key highlights:
Print-and-scan processing can increase the false match rate of face recognition systems by up to 5.64% for DiM morphs and 16.00% for StyleGAN2 morphs, compared to digital-only morphs.
Using the Fréchet Inception Distance (FID) metric, strictly print-scanned morph attacks performed on average 9.185% stronger than non-print-scanned digital morphs in terms of visual fidelity.
The paper examines four attack configurations that combine digital and print-scanned bona fide and morphed images, highlighting the vulnerabilities of face recognition systems to heterogeneous data.
Experiments were conducted on the FRLL, FERET, and FRGC datasets using three face recognition systems: ArcFace, ElasticFace, and AdaFace.
The results show that print-and-scan processing can effectively mask morphing artifacts, making it harder for face recognition systems to detect morphed images, especially when the reference data is also print-scanned.
Stats
Print-and-scan processing can increase the false match rate by up to 5.64% for DiM morphs and 16.00% for StyleGAN2 morphs.
Strictly print-scanned morph attacks performed on average 9.185% stronger than non-print-scanned digital morphs in terms of visual fidelity.
Quotes
"Print-and-scan processing can effectively mask morphing artifacts, making it harder for face recognition systems to detect morphed images, especially when the reference data is also print-scanned."
"The presence of inconsistent data when comparing potentially morphed images against bona fide images is an area lacking research."