Research On Participant Selection Mechanism Of Federated Learning Based On Attribute-based Encryption

Federated learning is a privacy protection paradigm for training machine learning models.It allows participants to train models locally and upload model parameters to a central server for aggregation.In this process,all sensitive data are stored locally,reducing the risk of privacy leakage.However,in real scenarios,the participants of federated learning usually present significant heterogeneity in terms of data distribution and hardware parameters.Therefore,random sampling of participants in each training round may not make full use of local updates of heterogeneous clients,resulting in low accuracy of the model,slow convergence rate and other problems;In addition,the lack of reasonable incentive mechanism will also weaken the willingness of participants to share the model.Therefore,how to select right participants to join the training and how to motivate participants properly has become a research focus in the scenario where data are Non independent and identically distributed.Based on attribute-based encryption technology and from the perspective of access control,this thesis proposes a bilinear pair-based synchronous federated learning participant selection and incentive mechanism and a pairing free hierarchical asynchronous federated learning participant selection mechanism to improve the performance of the federated learning model on the premise of ensuring security and privacy.Specifically,the main work of this thesis are as follows:1.In order to solve the problem of Non-independent and identically distribution of data in synchronous federated learning,a participant selection algorithm based on pairing-based attribute base encryption is proposed in this thesis.The central server of federated learning can set appropriate access policies before training,and only the participants who meet the requirements can successfully decrypt training flags and join the training process.At the same time,the contributions of the participants to the model are properly recorded in the blockchain for equitable distribution of training benefits.In this thesis,the IND-CPA security of the algorithm is rigorously proved,and the proposed participant selection strategy can improve the training accuracy of the federated learning model through a series of experiments.Compared with traditional random selection strategies and other participant selection schemes in the literature,the performance of the proposed strategy is improved.2.In order to solve the problem of device heterogeneity and stragglers,this thesis proposes a participant selection algorithm for hierarchical asynchronous federated learning based on pairing-free attribute based encryption,and presents a Cloud-Edge-End hierarchical federated learning framework.The aggregation server encrypts the global model using the access policy and distributes it to fog nodes for semi-decryption.Participants use unique private keys to further decrypt the global model and upload local model after a round of training using local data.The fog node synchronously aggregates the local updates of the subject partcipants,and the aggregation server asynchronously obtains the model of the edge node,and calculates the stale weight to correct the learning rate.In the same way,the IND-CCA security of the proposed scheme is further demonstrated.Experiments show significant advantages of the proposed scheme in terms of model performance and time overhead.