Research On Dynamic Asynchronous Federated Learning With Privacy Protection

Federated learning,as a newly proposed collaborative learning method,enables clients to perform machine learning training locally,protects local data from leaking,and breaks the data barrier between organizations or nodes.However,with the emergence of new application scenarios,such as vehicular ad hoc network,Internet of Things,etc.,the tra-ditional synchronous federated learning architecture with servers faces problems such as single points of failure,which can not meet the needs of group dynamics in some spe-cial application scenarios.This paper proposes a new serverless asynchronous federated learning scheme.We analyze the design goals of our scheme.Our design goal is to de-sign a privacy protection and asynchronous aggregation scheme suitable for serverless asynchronous federated learning,and solve the problem of data security and low-quality update.The main works of this paper come as follows:· Analyzes and summarizes the security threats and challenges of existing synchronous and asynchronous federated learning aggregation and privacy protection schemes.· We propose a dynamic asynchronous federated learning aggregation scheme based on dataset and model accuracy verification.Federated learning can effectively break data barriers,and can combine distributed clients for multi-party intelligent learn-ing while keeping data locally.However,the existing schemes based on federated learning do not consider the dynamics of the client in some new application scenar-ios(such as vehicular network,etc.).Serverless asynchronous federated learning may receive low-quality model updates during the aggregation process,which af-fects the accuracy of final model.Aiming at the aggregation issues,we analyse the design goals of the dynamic asynchronous federated learning aggregation scheme in detail.We designed an asynchronous aggregation scheme based on dataset and model accuracy verification.Using the methods of temporary aggregation and sim-ilarity comparison,we can identify and discard low-quality updates based on the verification results,to improve the performance of the final training model.This scheme can effectively meet the needs of asynchronous aggregation in a dynamic environment and avoid low-quality update attacks.Further,we use performance evaluation and experimental analysis to prove the effectiveness of our scheme.· We propose a privacy protection scheme for asynchronous federated learning.Asyn-chronous federated learning relies on the data interaction between a large number of clients,the exchange of model updates.During this process,the curious or dis-honest clients may obtain the models to acquire the data privacy of honest clients through vulnerable transmission channels.Therefore,the confidentiality and au-thentication of data need to be considered.On the other hand,in order to ensure the credibility of the data source,it is necessary to authenticate the clients in the group to prevent risks from unauthorized nodes.It is necessary to study the establishment of dynamic controllable confidentiality and authentication channel conforming to the characteristics of asynchronous federated learning.We propose to use contribution broadcast encryption combined with differential privacy and digital signature tech-nology to protect data privacy and resist inference attacks,reconstruction attacks and other related attacks.Clients who participate in asynchronous federated learn-ing at the same time usually change rapidly over time.For instance,the changes of members in the intelligent vehicle group will be very frequent.We expand the contribution broadcast encryption scheme to realize the dynamic control of group members and ensure the confidentiality and authentication of data interaction be-tween clients.We perform experimental simulation to analyze the feasibility and efficiency of our scheme.