StyleDemorpher: high-quality face demorphing via StyleGAN2’s latent space

Abstract Morphing attacks pose a serious threat to automated border control systems by allowing identity documents to be used by multiple individuals, undermining biometric security. To address this, we propose a novel face demorphing framework that leverages the latent space of StyleGAN2. At its core is ReStyle-ID, an encoder network optimized for identity preservation through improved loss functions and targeted training data, enabling accurate and identity-focused inversion. Combined with StyleDemorpher, a face demorphing network trained on a novel DemorphDB dataset with high-quality morph images that simulate realistic and challenging attack scenarios, the framework reconstructs high-resolution demorphed faces and generalizes well to unseen identities and morphing methods. Together, these components overcome key limitations of prior approaches, such as low resolution, poor robustness, and visual artifacts. This work offers a scalable and effective solution for face demorphing and contributes a comprehensive dataset and framework to support future research in biometric security.