MLSD-GAN -- Generating Strong High Quality Face Morphing Attacks using Latent Semantic Disentanglement

TL;DR

MLSD-GAN leverages latent semantic disentanglement in StyleGAN to generate realistic, high-quality morphs by splitting latents into identity and attribute components and applying spherical interpolation along a latent transfer direction derived from facial landmarks. The method uses a landmark-guided encoder (pSp) to produce $\mathbb{W}^+$ latents, with CodeFormer post-processing to boost perceptual quality. Evaluations on ArcFace and MagFace with FRGC-V2 data show the proposed morphs yield high vulnerability (G-MAP around $90$–$93\%$) compared with baselines, underscoring a significant security risk for current FRS. The work highlights the need to incorporate such advanced morphing attacks into Morph Attack Detection pipelines and informs defense strategies by clarifying the role of latent-space manipulation in morphing threats.

Abstract

Face-morphing attacks are a growing concern for biometric researchers, as they can be used to fool face recognition systems (FRS). These attacks can be generated at the image level (supervised) or representation level (unsupervised). Previous unsupervised morphing attacks have relied on generative adversarial networks (GANs). More recently, researchers have used linear interpolation of StyleGAN-encoded images to generate morphing attacks. In this paper, we propose a new method for generating high-quality morphing attacks using StyleGAN disentanglement. Our approach, called MLSD-GAN, spherically interpolates the disentangled latents to produce realistic and diverse morphing attacks. We evaluate the vulnerability of MLSD-GAN on two deep-learning-based FRS techniques. The results show that MLSD-GAN poses a significant threat to FRS, as it can generate morphing attacks that are highly effective at fooling these systems.

Figures (4)

• Figure 1: Block diagram of the proposed MLSD-GAN for generating high quality morphed face images
• Figure 2: Qualitative results of the proposed MLSD-GAN with existing methods a) Landmarks-I b) Landmarks-II, c) StyleGAN, d) MIPGAN-1 e) MIPGAN-II, f) Morrdiff g) Proposed method
• Figure 3: Spherical interpolation between subject-1 and subject-2 latents where $\alpha$ is the interpolation factor lies between 0 to 1 in our case $\alpha=0.5$ and $\theta$ is the angle subtended by the arc on the unit sphere.
• Figure 4: Box plots of PSNR values computed from different face morph generation methods