Research On Privacy-preserving Classification Method And Application Based On Federated Learning

With the continuous increase of data assets from enterprises,governments and private individuals,the demand for classification applications such as images in the field of machine learning is also increasing.In order to meet various practical needs,the idea of cloud service deployment in Machine Learning as a Service(MLAAS)has gradually become the mainstream.However,the large amount of data required to implement machine learning applications based on cloud services is often provided by multiple different parties,such as enterprises and governments.If the data is uploaded to the cloud server openly and transparently,it will be accompanied by serious data privacy security issues.In terms of protecting the data privacy of the participants,the idea of federated learning(FL)guarantees the security of its local data from the system level,but because the parameters required in the training model process are transmitted in plaintext on common channels,so it will leak data features and model information inevitably and indirectly.Even if the parameters are directly encrypted using cryptographic technology,the complexity of the cryptographic method will cause incalculable communication overhead between the participants and the cloud server.Therefore,how to greatly reduce the overhead between participants and cloud servers on the basis of protecting data privacy and security is very important.This thesis will focus on the distributed idea of federated learning to study more efficient privacy protection classification methods and application issues.Firstly,this thesis proposes a federated learning privacy-preserving classification scheme based on crowdsourcing aggregation.It crowdsources classification tasks to multiple edge participants and uses cloud computing to complete the whole process.However,instead of using the method of jointly training ideal models to obtain high-confidence classification results,we let the participants first train model based on limited local data and use the model to infer,and then we use mature algorithms to aggregate the inference results to obtain classification with higher accuracy.During the system implementation,we use homomorphic encryption to encrypt image data that requires machine learning inference;we also improve a crowdsourced federated learning classification algorithm,and implement the privacy-preserving computation of the entire system by introducing a dual-server mechanism.Through correctness analysis and security proof,this scheme ensures that the data query party will not disclose any private data,and also solves the privacy security problem.Secondly,this thesis designs and implements a federated learning privacy-preserving classification system based on crowdsourcing aggregation.It effectively reduces the communication overhead between the participants and the cloud server in the classification process and ensures the privacy and security of data.This thesis describes the general architecture and layered structure of the system,and carries out detailed design,requirement analysis and module design for the main modules.Finally,the performance test of the system is carried out,and the analysis shows the feasibility of the system.The security degree of privacy protection has been significantly improved,which is more suitable for application scenarios with higher privacy security requirements in real life.