Research And Application Of Communication Security And Efficiency In Chained Federated Learning

In recent years,federated learning has developed rapidly and has been more widely used in our practical applications.The issues of privacy security and communication efficiency in federated learning have also received more attention and research.Among them,chained structure and compression technology are important solutions in federated learning.By combining these two methods,we can improve the communication efficiency in federated learning training while ensuring the privacy and security of users.This paper firstly introduces the background and research status of federated learning,and analyzes some classical privacy protection algorithms in detail.Aiming at the phenomenon that the existing privacy protection methods and communication optimization methods can not take into account both privacy,computational complexity and model training accuracy in federated learning,relevant improvements are proposed.The main work is as follows:(1)The traditional privacy-preserving federated learning method faces the risk of privacy disclosure and model poisoning attack.To solve the above problems,the chained federated learning framework solves the gradient leakage problem,and uses the chained communication mechanism to make the masking gradient transmitted between participants.In order to improve the security of federated learning,an Adversarial-Aware model based on chained federated learning is proposed to prevent model poisoning attacks from damaging the training effect.In the Adversarial-Aware model,a malicious node identification method is designed.This method avoids the model poisoning attack by the value of user credibility and grouping.The Adversarial-Aware model is divided into groups according to the credibility of different users,which are divided into malicious attackers,suspects and ordinary users.In the chained training process,adjust the user’s credibility according to the training accuracy of the current group,gradually identify the malicious attacker from all clients,and do not allow the attacker to participate in the subsequent training again.In this paper,experiments on CIFAR-100 data set show that the model can effectively prevent the model degradation caused by model poisoning attack,improve the security of the training process of Federated learning,and ensure better training speed.(2)The traditional federated learning framework often has the problems of insufficient accuracy and low communication efficiency.Chained communication structure is introduced to defend against privacy disclosure attack in federated learning,but the large model parameters increase the communication cost,resulting in low communication efficiency.In order to effectively improve the efficiency of network communication,a communication efficiency optimization model based on chained federated learning,also known as chained gradient compression model,is proposed.The model performs gradient compression processing for user gradient parameters,sets the absolute value of the gradient ranking 15% as the threshold,prunes the gradient parameters by setting the gradient threshold,and transmits only the gradient value whose absolute value is greater than the threshold.In the process of training,the communication between the federated nodes is reduced and the transmission efficiency of all users’ content is improved.Experiments on MNIST and CIFAR-10 public data sets show that the model can effectively balance the privacy and communication efficiency in the training process.In order to evaluate the performance of the model,the gradient compression rate,privacy and convergence are analyzed respectively.The experimental results indicate that the accuracy of the model is close to the comparison framework,and the communication efficiency is much faster than the comparison scheme.(3)This paper presents the application and implementation of a federated secure communication framework in machine translation.In this paper,federated learning is applied to the field of machine translation,and combines with multi-user training model to improve the accuracy of machine translation and the security in the training process.The system combines the chained communication structure,attack perception model and chained gradient compression model.The chained structure is used to protect the transmitted gradient information and achieve the same federated training effect as Fed Avg while defending against gradient leakage attack.The Adversarial-Aware model is used to monitor the model poisoning attack during the training process,prevent the poor training effect and model degradation caused by the attack,and improve the security in the federated training process.The chained gradient compression model is used to compress the gradient parameters,and improve the efficiency of network communication.