SilverMask: Face Template Protection with Fine-Grained Noise Correction

Abstract

As face recognition systems (FRSs) become widely deployed in practice, the need to protect face templates has been highlighted for their security and privacy. Although one of the major benefits of FRSs is that they do not rely on any user-specific secret information, this property also imposes strong constraints on template protection schemes. The fuzzy commitment (FC) scheme is a promising tool for this setting, yet existing FC-based methods suffer from significant accuracy degradation compared to non-protected FRSs. To address this limitation, we present SilverMask, an FC-based face template protection scheme based on a novel real-valued error-correcting code (ECC), SilverCode. We observe that the accuracy degradation in prior FC-based methods stems from the insufficient error-correcting capacity of the underlying ECC to handle the high intra-class variation. SilverCode addresses this mismatch by its enhanced error-correcting capacity, which is derived from its explicit algebraic structure designed for real-valued face templates. Furthermore, we introduce GIC loss, a novel loss function that constrains the recognition model’s embedding space to encourage intra-class embeddings to align with decision boundaries of FC-based template protection schemes. These two techniques effectively bridge the gap between insufficient error-correcting capacity and high intra-class variance, thereby substantially reducing the accuracy loss observed in previous FC-based template protection schemes. To validate the effectiveness of these techniques, we conduct extensive experiments with four representative benchmark datasets (LFW, CFP, AgeDB, and IJB-C). Notably, SilverMask improves the TAR by 35.57% on the LFW dataset compared to the previous FC-based template protection baseline, while ensuring a 115-bit security level. To facilitate future research, we publicly release our source code on GitHub.