المفاهيم الأساسية
The proposed Diffusion-based face morphing attack generates high-quality morphed images that can effectively deceive state-of-the-art face recognition systems.
الملخص
The paper presents a novel face morphing attack that leverages Diffusion-based generative models to improve the visual fidelity and effectiveness of the generated morphed images. Key highlights:
- The Diffusion-based attack outperforms other state-of-the-art morphing attacks, including GAN-based and Landmark-based approaches, in terms of visual fidelity as measured by the Fréchet Inception Distance (FID).
- Extensive experiments show the Diffusion-based attack is highly effective at deceiving three leading face recognition systems (FaceNet, VGGFace2, ArcFace) across three different datasets (FRLL, FRGC, FERET).
- The Diffusion attack consistently ranks among the top performers in terms of the Attack Presentation Classification Error Rate (APCER) at specific Bona fide Presentation Classification Error Rate (BPCER) values.
- A novel metric, Mated Morphed Presentation Match Rate (MMPMR), is introduced to measure the relative strength of different morphing attacks. The Diffusion attack outperforms other attacks on this metric on average.
- The impact of the face recognition system's pre-processing pipeline, particularly the image cropping, is explored. Tighter cropping makes the system more resilient against attacks with visual artifacts outside the core face region.
الإحصائيات
The proposed Diffusion-based morphing attack achieves an FID of 42.63, 64.16, and 50.45 on the FRLL, FRGC, and FERET datasets respectively, outperforming other state-of-the-art attacks.
On the FRLL dataset, the Diffusion attack achieves an APCER of 8.83%, 2.68%, and 3.33% at a BPCER of 0.1% for the FaceNet, VGGFace2, and ArcFace face recognition systems respectively.
On the FRGC dataset, the Diffusion attack achieves an APCER of 91.73%, 93.71%, and 40.6% at a BPCER of 0.1% for the FaceNet, VGGFace2, and ArcFace face recognition systems respectively.
On the FERET dataset, the Diffusion attack achieves an APCER of 24.04%, 80.9%, and 9.69% at a BPCER of 0.1% for the FaceNet, VGGFace2, and ArcFace face recognition systems respectively.
اقتباسات
"The proposed Diffusion-based face morphing attack generates high-quality morphed images that can effectively deceive state-of-the-art face recognition systems."
"The Diffusion attack consistently ranks among the top performers in terms of the Attack Presentation Classification Error Rate (APCER) at specific Bona fide Presentation Classification Error Rate (BPCER) values."
"Tighter cropping makes the face recognition system more resilient against attacks with visual artifacts outside the core face region."