Efficient And Privacy-preserving Federated Learning With Irregular Users

Federated Learning(FL)is a deep learning algorithm,which enables users to learn a global model by exchanging local updates without disclosing the private dataset.In order to further protect the privacy of users,many privacy protection methods are applied in FL.However,these works rarely consider the fundamental problem that the updates shared by a part of users(called irregular users)is low quality.In order to solve this problem,this paper proposes two secure aggregation schemes for the FL of privacy protection for irregular users for honest clients and malicious clients respectively.Specifically,the main contributions of this thesis are as follows:For honest clients,this paper proposes a FL framework for privacy protection,TRUST-IU,which excludes irregular users.TRUST-IU calculates the weight of each user for model aggregation by using cosine measures.TRUST-IU reduces the negative impact of irregular users,and ensure that the global model comes primarily from high-quality data contributions.TRUST-IU employs a double-mask technique to ensure that the server only learns the aggregated results of the user’s local gradients.In addition,the scheme is robust to exiting users during the whole training process.For malicious clients,thesis proposes a FL system with input validation to exclude irregular users-UNTRUST-IU.UNTRUST-IU identifies malicious users and irregular users by combining Euclidean distance and Chebyshev distance,and eliminates their negative effects.In terms of privacy protection,UNTRUST-IU employs a more effective secure aggregation scheme and extends it by enabling input validation.In this validation,the server can check whether the input of the client meets the required constraints,so as to ensure that the server only learns the aggregation results of honest and high-quality local gradients.In addition,the scheme is robust to exiting users during the whole training process.