Research On Data Aggregation Security Protocol In Horizontal Federated Learning Environment

With the rapid development of artificial intelligence,more and more fields are focusing on this.At the same time,some projects have been implemented.However,the development of artificial intelligence and the accuracy of models rely on a large amount of data.In practice,most companies have small and low-quality data except for a few companies that can meet the requirements of data amounts.They are unable to construct an accurate model.Especially,domestic and foreign institutions have formulated relevant laws and regulations to restrict the flow of data and strengthen the requirements for data privacy protection.In order to solve the “data island” problem,federated learning is proposed.At present,there are many commercial application cases of horizontal federated learning,which indirectly expand the amount of data for model training.However,horizontal federated learning is still in the primary stage.In particular,there are still many technical challenges in data privacy protection.However authenticated key agreement protocol is an important technology to ensure secure transmission and reliable communication between communication entities.This paper focuses on the security of the horizontal federated learning data aggregation protocol and selects the application scenario of a single server.The main research contents are as follows:(1)This paper designs a two-party authentication certificateless key agreement protocol using an identity-based cryptosystem without bilinear pairing in a single-server horizontal federated learning environment.In this protocol,the participants do not rely on the key generation center,and they will use public parameters to calculate temporary keys and long-term keys.At the same time,they use fresh random numbers during interaction process to ensure the freshness of the session keys.The eCK model is introduced into the security proof of the protocol.While satisfying the basic security,the protocol also has security attributes such as forward security,resistance to unknown key sharing attacks,key leakage attacks,and two-way authentication.It avoids unnecessary computational overhead,which is authentication firstly and then computes session key.Therefore,it is comprehensively analyzed that this protocol has complete security properties,lower computing and communication costs,and it is suitable for a single-server horizontal federated learning environment.(2)This paper focuses on analyzing the security requirements of the horizontal federated learning data aggregation protocol,it improves the existing training model and designs a new aggregation method--Federated Weighted Average Aggregation(Fed WAvg).It considers the difference of the model in user's data quality,and computes the parameter weights according to the difference between the local model and the global one.Then these parameter weights are encrypted with the homomorphic encryption algorithm and they are upload.The aggregation server selects partial users to participate in this aggregation according to the upload situation of each user,then the server computes the global model parameters in the ciphertext state.Compared with the existing federated average aggregation training architecture and unencrypted one,the data aggregation protocol proposed in this paper can improve computational efficiency and accuracy through experimental simulations.Meanwhile it protects data privacy and ensures the security of data aggregation protocol.