GMoE-FTI: A Universal Facial Template Inversion Method Adopting Genuine Mixture-of-Experts Mechanism

Computer Science > Computer Vision and Pattern Recognition arXiv:2610.33055 (cs) [Submitted on 19 Oct 2026] Title: GMoE-FTI: A Universal Facial Template Inversion Method Adopting Genuine Mixture-of-Experts Mechanism.

计算机科学 > 计算机视觉与模式识别 arXiv:2610.33055 (cs) [提交于 2026 年 10 月 19 日] 标题：GMoE-FTI：一种采用真实混合专家机制的通用人脸模板反演方法。

Abstract: Face recognition systems have been ubiquitously deployed across daily scenarios. To protect user privacy, mainstream systems store facial templates rather than raw face images for identity verification. However, facial template inversion (FTI) technologies can reconstruct faces from templates, which endangers biometric privacy.

摘要：人脸识别系统已在日常场景中得到广泛部署。为了保护用户隐私，主流系统存储人脸模板而非原始人脸图像以进行身份验证。然而，人脸模板反演（FTI）技术可以从模板中重建人脸，这危及了生物识别隐私。

Existing FTI methods only fit single type of templates, failing to reconstruct identity-consistent faces from unseen types. To tackle this limitation, we survey mainstream template extractors which can be classified into five categories according the template distribution: unit-sphere, angle-linear, strict unit-sphere, cone-shaped, and uniform hypersphere.

现有的 FTI 方法仅适用于单一类型的模板，无法从未知类型中重建身份一致的人脸。为了解决这一局限性，我们调查了主流模板提取器，根据模板分布可将其分为五类：单位球、角度线性、严格单位球、锥形和均匀超球。

Inspired by this observation, this paper proposes GMoE-FTI, a universal FTI framework powered by a genuine MoE mechanism that differs from load-balanced multi-branch alternatives. It follows a three-stage training strategy: (1) independently train five mapping experts (MEs) and a matching module (MM); (2) conduct collaborative inference by weighting each ME’s output with MM’s output, so as to generate unified latent codes compatible with pre-trained StyleGAN3; (3) perform end-to-end joint fine-tuning over all modules to ensure identity consistency and boost cross-expert coordination.

受此观察启发，本文提出了 GMoE-FTI，这是一个由真实 MoE 机制驱动的通用 FTI 框架，它不同于负载均衡的多分支替代方案。它遵循三阶段训练策略：（1）独立训练五个映射专家（ME）和一个匹配模块（MM）；（2）通过 MM 的输出对每个 ME 的输出进行加权，进行协同推理，从而生成与预训练 StyleGAN3 兼容的统一潜在代码；（3）对所有模块进行端到端的联合微调，以确保身份一致性并增强跨专家协作。

Extensive evaluations on the LFW benchmark validate the superiority of GMoE-FTI against SOTA methods. Evaluated on seven prevailing unseen template types including AdaFace, CosFace, CurricularFace, ElasticFace, SphereFace2, SphereFace-R, and TransFace, GMoE-FTI achieves an average 40% TAR improvement, with average FAPD and FAPC each reduced by 2%, verifying its universal generalization across heterogeneous templates.

在 LFW 基准测试上的广泛评估验证了 GMoE-FTI 相较于现有最先进（SOTA）方法的优越性。在包括 AdaFace、CosFace、CurricularFace、ElasticFace、SphereFace2、SphereFace-R 和 TransFace 在内的七种主流未知模板类型上进行评估，GMoE-FTI 实现了平均 40% 的 TAR 提升，平均 FAPD 和 FAPC 各降低了 2%，验证了其在异构模板上的通用泛化能力。