结合双随机相位加密和四元数格拉斯曼平均网络的彩色人脸隐私保护识别

Privacy-Protected Color Face Recognition Using Double Random Phase Encoding and Quaternion Grassmann Average Networks

为保护彩色人脸图像的隐私并提高识别准确率,提出一种结合双随机相位加密和四元数格拉斯曼平均网络的彩色人脸隐私保护识别算法.首先将彩色人脸图像的颜色分量编码为纯四元数矩阵,使用双随机相位加密和四元数Gyrator变换进行加密.为了隐藏原始彩色人脸图像的内容,通过随机二值幅度掩模选取部分密文进行解密,得到不可见的人脸图像,然后使用四元数格拉斯曼平均网络提取人脸的特征向量和线性支持向量机进行识别.当人脸特征模板泄露时,可以重新生成随机二值幅度矩阵并构建新的特征模板替换原始的特征模板,满足可撤销性,从而保证原始人脸图像的安全性.在Aberdeen, Georgia Tech, Visible Light和YouTube Makeup 4个数据集上与当前3种人脸隐私保护识别算法进行比较,实验结果表明该算法能够有效地提高识别率,而且对数据集的变化具有较好的鲁棒性.

To protect the privacy and improve the recognition accuracy,a method of color face privacy protection recognition based on double random phase encoding and quaternion Grassmann average networks is proposed.Firstly,three color components of each color face image are encoded into a pure quaternion matrix.Then the quaternion value matrix is performed double random phase encoding in quaternion gyrator domain.In order to conceal the content of color facial image,only a small part of encrypted data is randomly selected by using random binary amplitude mask and preserved for decryption.For the invisible decrypted face images,quaternion Grassmann average networks is employed to extract features and recognition rate is calculated by linear support vector machine.In case that the face template is leaked,alternative one can be reproduced by new random binary amplitude matrix,which makes face template cancelable and keeps original face images in safe.The four datasets were used including Aberdeen,Georgia Tech,Visible Light and YouTube Makeup and the recognition rates are compared with other three face image privacy protection methods,the experimental results have demonstrated that the proposed method can effectively improve the recognition rates and it shows better robustness to the change of dataset.