BinaryFace:基于深层卷积神经网络的人脸模板保护模型

随着生物识别技术的广泛应用，人们越来越担心生物模板信息的安全性和隐私性。为此人们提出很多生物模板信息的保护算法，但其一般需要牺牲可识别性来换取高安全性。为了在保证高安全性的同时尽可能提高可识别性，本文提出一种新的由特征转换和生物加密组成的二阶段人脸模板保护方案。在特征转换阶段，基于VGGFace提出一种新的基于卷积神经网络的BinaryFace网络，通过设计新的随机正交映射矩阵、量化损失函数和最大熵损失函数实现人脸模板的二进制转换。同时为了减少网络参数，设计新的深度可分离瓶颈卷积层，BinaryFace相比VGGFace在参数和浮点数（Flops）上分别减少约75%和约35%。在生物加密阶段，将人脸二进制模板转换中随机正交映射生成的纠错码输入模糊承诺方案，生成加密的人脸模板并存储到数据库中。在验证阶段，通过相同的流程恢复出纠错码，并与原始纠错码进行哈希校验得到最终的匹配结果。在评测阶段，本文提出的方法在CMU-PIE、FEI、Color FERET等3个数据集上，相比之前的工作在GAR上有约6.5%的提升，同时将EER降低了约4倍。

BinaryFace: the Model of Face Template Protection based on CNN

With the widespread use of biometrics, there is growing concern about the security and privacy of biometric template information. So many protection algorithms based on biological template information have been proposed, but they generally need to sacrifice recognizability in exchange for high security. In order to ensure high security while maximizing recognizability, this paper proposes a new two-stage face template protection scheme consisting of feature conversion and bio-encryption. In the feature conversion stage, a new BinaryFace network based on convolutional neural networks was proposed based on VGGFace. Binary conversion of face templates was achieved by designing a new random orthogonal mapping matrix, quantization loss function, and maximum entropy loss function. At the same time, in order to reduce the network parameters, a new deep separable bottleneck convolution layer was designed. Compared with VGGFace, BinaryFace reduced parameters and floating point numbers (Flops) by about 75% and 35%, respectively. In the bio-encryption phase, the key generated by the random orthogonal mapping in the binary binary template conversion is input into the fuzzy commitment scheme, and the encrypted face template is generated and stored in the database. In the verification phase, the key is recovered through the same process and compared with the original key to obtain the final matching score. At the evaluation stage, the method proposed in this paper has about 6.5% improvement in GAR on the three datasets of CMU-PIE, FEI, Color FERET, etc., while reducing EER by about 4 times.