Boutros Fadi ; Damer Naser ; Kuijper Arjan

Abstract

Deep learning-based face recognition models follow the common trend in deepneural networks by utilizing full-precision floating-point networks with highcomputational costs. Deploying such networks in use-cases constrained bycomputational requirements is often infeasible due to the large memory requiredby the full-precision model. Previous compact face recognition approachesproposed to design special compact architectures and train them from scratchusing real training data, which may not be available in a real-world scenariodue to privacy concerns. We present in this work the QuantFace solution basedon low-bit precision format model quantization. QuantFace reduces the requiredcomputational cost of the existing face recognition models without the need fordesigning a particular architecture or accessing real training data. QuantFaceintroduces privacy-friendly synthetic face data to the quantization process tomitigate potential privacy concerns and issues related to the accessibility toreal training data. Through extensive evaluation experiments on sevenbenchmarks and four network architectures, we demonstrate that QuantFace cansuccessfully reduce the model size up to 5x while maintaining, to a largedegree, the verification performance of the full-precision model withoutaccessing real training datasets.