中国科学院自动化研究所机构知识库

       随着深度学习技术的飞速发展，人脸识别技术得到了巨大的提升。其作为目 前计算机视觉领域中发展较快的一门技术，已被应用于生活中的各个方面。然而 基于深度学习的人脸识别模型往往需要一个庞大的人脸数据集用于训练才能获 得高识别率。但是，人脸属于用户的个人隐私，在未经本人同意之前不得收集和 利用。因此，大规模的人脸数据集收集的过程有可能会涉及侵犯个人隐私的问 题。为了解决隐私问题，联邦学习是一个好的解决方案。然而由于人脸识别模型 中的类别向量包含隐私信息，因此各个客户端之间无法共享类别向量，这使得当 客户端仅有一类人脸数据时，传统的联邦学习方法无法直接应用于人脸识别模 型的训练过程。同时，由于人脸识别模型一般较大，而在联邦学习框架中，客户 端与服务器端之间需要频繁通信，因此会给通信网络造成巨大的压力，导致训练 时间大大延长或者训练过程直接中断。因此，本文针对基于联邦学习的人脸识别 模型训练过程中所遇到上述问题进行了研究。

       首先，本文提出了一种基于等效类别向量的联邦人脸识别方法。其中，等效 类别向量能够保护客户端类别向量隐私同时使各个客户端类别向量分离以实现 隐私保护和提升模型性能的效果。该框架同时与各种人脸识别损失兼容。实验结 果表明本文所提出的方法在各个测试数据集上的表现均要优于现存的联邦学习 方法。

       解决了联邦人脸识别模型训练问题之后，需要解决训练时间过长的问题。因此，本文对于提高联邦人脸识别训练过程中的通信效率进行了研究。本文结合前 人的梯度量化工作，提出了全自适应梯度量化算法。该算法能够在训练过程中实 现量化层级和量化区间端点同时自适应地改变，以在每一轮获得接近最优量化 的梯度量化方案。最后的实验结果表明，该算法相较于现有的梯度量化算法而 言，其所需的通信量更低，同时精度损失最少。

       综上所述，本文所提出的两种方法在一定程度上缓解了基于联邦学习的人 脸识别方法所遇到的问题。

       With the rapid development of deep learning technology, face recognition has been greatly improved. As a rapidly developing technology in the field of computer vision, it has been applied to many aspects of life. However, face recognition models based on deep learning often need a huge face dataset for training in order to obtain high accuracy. However, face images belong to the user’s personal privacy and cannot be collected and used without user’s consent. Therefore, the process of collecting large-scale face dataset may involve the violation of personal privacy. In order to solve the privacy problem, federated learning is a good solution. However, because the class embeddings in the face recognition model contain privacy information, the class embeddings cannot be shared among clients, which makes the classic federated learning algorithms can not be directly applied to the training process of the face recognition model when each client has access to only one class of face data. At the same time, because the face recognition model is generally large, and in the federal learning framework, frequent communication is required between the clients and the server, thus, it will cause great pressure on the communication network, resulting in a significant extension of training time or direct interruption of the training process. Therefore, this thesis studies the above problems in the process of training face recognition model based on federated learning.

       Firstly, this thesis proposes a federated face recognition method based on Equivalent Class Embedding. The Equivalent Class Embedding can protect the privacy of the clients’ class embeddings and separate each client’s class embedding to achieve the effect of privacy protection and improving the performance of the model. At the same time, the framework is compatible with various face recognition loss. The experiment results show that the performance of the proposed method is better than the existing federated learning methods in all test benchmarks.

       After solving the training problem of federated face recognition model, the problem that the training time is too long should be solved. Therefore, this thesis studies how to improve the communication efficiency in the training process of federated face recognition. This thesis proposes a method, AllAdaQuant, based on the previous gradient quantization works. The algorithm can adaptively change the quantization level and quantization interval endpoint at the same time in the training process, so as to obtain a gradient quantization scheme close to the optimal quantization in each round. Finally, the experiment results show that compared with the existing gradient quantization methods, our method requires less communication and obtain the least precision loss.

       To sum up, the two methods proposed in this thesis alleviate the problems of face recognition method based on federated learning to a certain extent.