Data selection method in federated learning scene

Abstract

A data selection method in a federated learning scene comprises the steps of filtering out users and data related to tasks, user selection before training, user and data selection in the training process and model training. According to the invention, a vector sketch and a random response mechanism are adopted, a user selection strategy is efficient, and privacy protection is achieved; meanwhile,the server side log information is adopted to dynamically select the user; data is selected based on the gradient upper bound value, and the influence of error data on the gradient is considered, so that the data selection strategy is efficient and accurate.

Description

technical field [0001] The data selection method in the federated learning scene involved in the present invention belongs to the field of data analysis and data quality evaluation. Background technique [0002] How to obtain large amounts of high-quality datasets has become a common bottleneck for many machine learning models and AI applications. This is not only because collecting and labeling large numbers of samples is expensive, but also because privacy concerns hinder data sharing in many fields, such as medicine and economics. The emergence of federated learning makes it possible for end users to jointly train network models using local data. In the process of federated learning, the quality of the user's local data affects the performance of the global model, and low-quality data (eg, wrongly labeled data, non-uniformly distributed data) will seriously hinder the global model from achieving good results. [0003] The present invention aims to select a group of high...

AI Technical Summary

Problems solved by technology

But they cannot be directly used in federated learning: 1) Existing methods require direct access to all training samples, while in federated systems, data cannot be directly accessed by third parties

2) Computing the importance of each sample directly creates an unacceptable overhead for actors with limited resources

3) Existing methods do not consider the impact of non-IID or wrong samples on the sample selection strategy, and may give higher importance to wrong samples, thereby reducing model performance

Images