| Federated learning is a popular distributed deep learning method,which is used in various industries to solve practical problems.Split learning is a new distributed deep learning method.Like federated learning,it can achieve complete model training without sharing the local data of the client.Split learning separates a complete neural network model and runs it on the client and server respectively,so it has better performance than federated learning in model privacy,and can also adapt to the client in resource constrained scenarios.Another feature of split learning is that the client running order is asynchronous serial,which results in the running speed of split learning is slower than that of federated learning.Some scholars have tried to apply federated learning related algorithms to split learning to solve the problem of split learning efficiency,but most of their solutions will bring privacy security problems to split learning.In terms of model privacy security of split learning,many scholars have proposed many research schemes,but split learning may have multiple or direct or indirect privacy problems in the training process,and any one of them may cause serious privacy leakage.In view of the above situation,this paper studies the model efficiency and privacy of split learning,and proposes an effective solution.In terms of model efficiency,this paper proposes a new multi server parallel split learning model,which divides the local data of each client into multiple copies.The model is trained based on multiple servers.After each client completes the model training with one copy of data,it can transmit the model to the back to complete the model update of the client.While the latter client is running,The client can use the local residual data to continue model training,so as to achieve model parallelism.Because each client only transmits the client model to the next client,the asynchronous serial update of the whole client is maintained.Experiments show that the method proposed in this paper has a greater degree of efficiency improvement than traditional split learning while ensuring the accuracy of the model.In terms of model privacy security,this paper considers multiple privacy problems that may exist in the training process of split learning,and uses differential privacy laplacian mechanism and exponential mechanism to avoid the direct exposure of the output data and data labels of the cut layer.In addition,a malicious server detection algorithm is designed by comparing data between multiple servers to find possible malicious servers,so as to avoid experimental verification,the method proposed in this paper has certain effectiveness. |
