Research On Privacy Preserving Machine Learning Technology Based On Federated Learning

With the rapid development of machine learning technology,machine learning has a great impact on social development.Machine learning has been applied to various aspects such as living,producing,and social working.Although machine learning brings many benefits to social development,it also brings the risk of privacy leakage.In recent years,researchers pay more and more attention to privacy protection.Privacypreserving machine learning tasks have gradually become a research hotspot.Federated learning technology is a popular technology in privacypreserving machine learning.It does not need to upload data,but only needs to upload training parameters.However,there are still some problems in the current federated learning technology.For example,using federated learning to optimize large models in the field of natural language processing is inefficient,especially in horizontal federated learning.In order to solve the above problems,we propose a federated learning framework based on memetic algorithm.The framework includes a model aggregation method based on the memetic algorithm,and a novel local search step.In order to solve the problems of network congestion and reduce training costs,we propose several local search strategies.In addition,we propose a new model aggregation algorithm based on memetic algorithm.Experiments show that the federated learning framework based on memetic algorithm can generate global models in fewer rounds,while the generated model has better performance on the test set.Furthermore,we develop a training evaluation system.The system can implement federated learning task training among multiple parties,and can also quantitatively evaluate the contributions of these parties.Serveral commonly used federated learning algorithm and the framework proposed in our paper are embedded in this system.We also added a contribution evaluation system based on Shapley value in the system.The organizers of the federated learning task can view the contributions of each parties in the system.The system is equipped with a visual interactive interface,which is convenient for users to perform operations such as system settings,realtime monitoring and viewing of training results.