Research And Implementation Of Distributed Machine Learning Open Service Platform For Big Data Credit Service

The big data credit industry uses massive financial data,combined with big data and artificial intelligence technology,to effectively assist private financial activities and national economic decision-making.However,business data in the big data credit industry faces the problem of isolated data island,and only distributed modeling for data privacy protection can be performed using federated learning.Federated learning is mainly for large-scale mobile training,which is different from smallscale server training such as big data credit reporting,which makes federated learning face many challenges in the process of landing.These include the Byzantine failure of federated learning,the problem of inefficient sampling,and the lack of a dedicated open platform.Based on the above problems,this paper has completed the following three points of work:1)Aiming at the problem that aggregation algorithms are vulnerable to Byzantine attacks in small-scale federated learning,a model aggregation algorithm based on historical backtracking is proposed.Exclusion of Byzantine models by existing Byzantine robust algorithms may occur after the attack hits and cannot eradicate the model poisoning caused by the attack.Therefore,our method uses the idea of backtracking,not only excludes the model parameters of the node during aggregation,but also excludes all the results of the node in recent rounds of iterations,so that the attacked model can quickly find the correct gradient descent direction.Experiments show that the algorithm is better than the traditional FedAvg algorithm and the Byzantine robust Krum algorithm under various conditions,which improves the training efficiency and reduces the communication overhead.2)Aiming at the problem that Byzantine nodes continue to threaten training tasks in small-scale federated learning scenarios,a federated learning model training system with node state management is designed and implemented.A state management module is proposed to divide nodes into multiple states to prevent Byzantine nodes from participating in training tasks;supplemented by a sandbox system to monitor the current state of nodes;a random sampling algorithm based on dynamic weights is proposed to reduce the monitoring process overhead.The system can not only prevent the occurrence of model poisoning,but also ensure that effective information is not lost.3)In view of the problem of high barriers to cooperation in the big data credit,an open service platform that supports various types of services has been built.Support the access,display,and use process of data,AI,and big data credit services,and design related data tables;registered AI services based on the above systems.It not only supports the joint modeling function with data privacy protection,but also promotes rich business cooperation between big data credit companies.Above all,this paper proposes an innovative design to eradicate and prevent Byzantine failure in the algorithm and system of federated learning,to improve the robustness of joint modeling;build an open service platform for big data credit services to reduce this problem.The research and development difficulties and cooperation barriers of related technologies in the field are of great practical significance.