Internet of vehicles federal learning hierarchical knowledge security migration method based on gradient memory

Abstract

The invention relates to an Internet of Vehicles federal learning hierarchical knowledge security migration method based on gradient memory, and the method comprises the following steps: 1, carrying out the clustering of a plurality of clients through employing a hierarchical clustering algorithm, obtaining a plurality of independent and identically distributed clusters, and merging a plurality of clients with heterogeneous data into the independent and identically distributed clusters; 2, establishing an Internet of Vehicles federated learning model based on a hierarchical cluster architecture; and step 3, performing knowledge migration among different clusters by adopting a knowledge migration federated learning algorithm based on gradient memory so as to alleviate the problem of disastrous forgetting of knowledge migration in the hierarchical cluster architecture. The method has the advantages that the problem of disastrous forgetting is relieved, and the model convergence speed and the model precision are effectively improved.

Description

technical field [0001] The invention relates to the technical field of federated learning of the Internet of Vehicles, in particular to a method for safely transferring layered knowledge of the federated learning of the Internet of Vehicles based on gradient memory. Background technique [0002] With the continuous growth of data volume, the improvement of computing hardware performance and the development of deep neural network, Internet of Vehicles, autonomous driving, etc. have made great progress in recent years, most artificial intelligence solutions are centralized, and users collect them All data transmitted to the central data server or the cloud, but this will bring privacy issues, delay and bandwidth limitations, in contrast, distributed architecture is a more privacy-preserving and efficient choice, federated learning (FL) is a An emerging distributed machine learning model, which allows all parties involved in learning to conduct cooperative training under the co...

AI Technical Summary

Problems solved by technology

[0003] In order to simplify the research conditions, it is usually considered that the data follow the assumption of independent and identical distribution, but in the real world, customer nodes often generate and collect data in a non-independent and identical distribution manner, which does not meet the commonly used assumption of independent and identical distribution, and in different The model convergence analysis of structural data modeling and training process poses challenges. When the original weighted aggregation algorithm such as federated aggregation is used for non-independent and identically distributed data, the global learning model converges to a stagnation point that is inconsistent with the real objective function. Thereby resulting in low model precision, in order to solve these problems and improve the convergence efficiency, the present invention adopts the hierarchical clustering algorithm to convert the non-IID data into multiple IID data clusters according to the similarity between the data, and the same data distribution The client nodes will be merged into a cluster and assigned specific tasks, each small cluster represents a small-scale federated system, such a large-scale independent and identically distributed federated learning system is divided into multiple small independent and identically distributed systems, however The different clusters are separated from each other and do not allow beneficial knowledge transfer between them

Images