目的 近年来,随着人脸识别认证技术的发展及逐渐普及,大量人脸照片存放在第三方服务器上的现象十分普遍,如何对人脸进行隐私保护这个问题变得十分突出。 方法 本文首先对人脸图像进行预处理,然后采用Arnold变换对人脸关键部位进行分块随机置乱,将得到的结果图输入到深度卷积神经网络中。为了解决人脸照片在分块置乱时由于本身拍照角度的原因导致的分块不均等因素,在预处理时根据人眼进行特性点定位,再据此进行对齐处理,使得预处理后的照片人眼处于同一水平线。针对人脸隐私保护及加扰置乱后图像的识别,本文提出了基于分块随机加扰的深度卷积神经网络,不包含附加层,该模型网络结构是由4个卷积层、3个池化层、1全连接层和1个Softmax回归层组成。服务器端通过深度神经网络模型直接对加扰后人脸图像进行验证识别。 结果 该算法使服务器端全程不存储原始人脸模板,实现了对原始人脸图像的有效加扰保护。实验采用该深度卷积神经网络对处理过后的ORL人脸库进行识别,最终识别准确率达到97.62%。同时通过多组对比实验,验证了本文所提出的方法的有效性。 结论 与其他文献中手工提取特征并利用决策树和随机森林进行训练识别的方法相比,本文所提方法减少了人工提取特征的工作量,并且仍然具有更高的识别率。
Recognition of face privacy protection using deep convolutional neural network
Zhang Jianwu,Shen Wei,Wu Zhendong(School of Communication Engineering,Hangzhou Dianzi University;School of Cyberspace,Hangzhou Dianzi University)
Objective In recent years, with the development and popularization of face recognition authentication technology, the phenomenon that a large number of face photos are stored on third-party servers is very common. Faces are also relatively open features, and many people post selfies on various social platforms. So the question of how to protect the privacy of the face becomes very prominent. Method In order to solve the problem of unevenness of the face due to the angle of the camera when the face photo is scrambled, this paper first preprocesses the face image: firstly, detecting the face whether is included in the image, if it exists, then find the border containing the complete face. Next you need to locate the key points such as the nose and eyes, align the face images according to these key point positions, and normalize to the same size according to the key mechanism of vision, that is, the human eye always sees the center of the photo first, and then gradually goes to the last four corners. After that, the key parts of the face (eyes, ears, mouth, nose) is scrambled blocked random times by Arnold transform. Secondly, aiming at face privacy protection and image recognition after scrambling, this paper proposes a deep convolutional neural network based on block random scrambling, which does not include an additional layer. The network structure of the model is composed of four convolutional layers, three pooling layers, one fully connected layer and a softmax regression layer. The convolution kernel sizes of the four convolutional layers are 6*6, 3*3, 3*3, 2*2, respectively. In the training phase, the preprocessed samples are divided into training sets and test sets. At the beginning of training, the convolution kernel parameters are randomly initialized to a small value, and small random numbers are used to ensure that the network does not enter saturated state due to excessive weights. The training process is divided into the forward propagation and backward propagation phases. After the input passes through multiple convolutional layers and pooling layers, it is transferred to the output layer. In the process, the input is actually multiplied by each layer of weight matrix, and the calculation is performed in turn to obtain the output result. The difference between the actual output and the ideal output is calculated in the backward propagation phase, and the weight is adjusted in reverse according to the minimization error method. The server side directly verifies and recognizes the scrambled face image through the deep neural network model. In order to further improve security, before the transmission or storage on the server, the pre-processed and randomized scrambled images are encrypted and the key is saved. At this time, the color histogram of the image will tend to be a straight line. When it is necessary to identify, if it has a legal key, it can be correctly restored to the previous state to perform the identification operation. Result This algorithm enables the server to not store the original face template throughout the entire process, thus achieving effective scrambling protection of the original face image. The experiment uses this deep convolutional neural network to identify the ORL face database, and the final recognition accuracy rate reaches 97.62%. At the same time, the effectiveness of the proposed method is verified by multiple sets of comparative experiments. Conclusion Compared with other methods of manually extracting features and using the decision tree and random forest for training recognition in other literatures, the proposed method reduces the workload of manual extraction features and still has a higher recognition rate.