LADIMO: Face Morph Generation through Biometric Template Inversion with Latent Diffusion

TL;DR

Face morphing threatens biometric verification by enabling a single morphed image to validate against multiple identities. LADIMO leverages a Latent Diffusion Model conditioned on biometric templates to invert MagFace embeddings and perform representational-level morphing via SLERP, starting from stochastic latent samples $z_T \sim \mathcal{N}(0,1)$. Across FRGCv2-based experiments and four state-of-the-art FRS, LADIMO delivers high-fidelity morphs and superior Morph Attack Potential (MAP) compared to MIPGAN-II, with robustness to unseen systems. The approach also introduces stochastic morph variation to generate diverse morph variants from a single image pair, highlighting implications for both morph-detection training and biometric security.

Abstract

Face morphing attacks pose a severe security threat to face recognition systems, enabling the morphed face image to be verified against multiple identities. To detect such manipulated images, the development of new face morphing methods becomes essential to increase the diversity of training datasets used for face morph detection. In this study, we present a representation-level face morphing approach, namely LADIMO, that performs morphing on two face recognition embeddings. Specifically, we train a Latent Diffusion Model to invert a biometric template - thus reconstructing the face image from an FRS latent representation. Our subsequent vulnerability analysis demonstrates the high morph attack potential in comparison to MIPGAN-II, an established GAN-based face morphing approach. Finally, we exploit the stochastic LADMIO model design in combination with our identity conditioning mechanism to create unlimited morphing attacks from a single face morph image pair. We show that each face morph variant has an individual attack success rate, enabling us to maximize the morph attack potential by applying a simple re-sampling strategy. Code and pre-trained models available here: https://github.com/dasec/LADIMO

Figures (9)

• Figure 1: Conceptual overview on LADIMO face morph generation, including a) the biometric template encoding meng2021magface, b) spherical linear interpolation shoemake1985animating, and c) LDM-based biometric template inversion results after t denoising steps.
• Figure 2: LADIMO training architecture overview, adopting the default hyperparameter settings from Rombach-LDM-CVPR-2022 and utilizing their pre-trained perceptual encoder and decoder. We customize the conditioning mechanism for biometric template inversion, learning to reconstruct face images from their corresponding MagFace meng2021magface embeddings.
• Figure 3: Example reference (top) and probe (bottom) samples from our FRGCv2 phillips2005overview subset.
• Figure 4: Example face morph images generated with LADIMO in comparison to MIPGAN-II Zhang-MIPGAN-TBIOM-2021.
• Figure 5: LADIMO Morph Attack Potential ISO-IEC-20059 in comparison to MIPGAN-II Zhang-MIPGAN-TBIOM-2021, assessed with three verification attempts and four FRSs: Deng-ArcFace-CVPR-2019Kim-AdaFace-CVPR-2022Huang-Curricularface-CVPR-2020meng2021magface with a fixed FMR at 0.1%.