Efficient contribution evaluation method in federated learning scene

Abstract

The invention discloses an efficient contribution evaluation method in a federated learning scene. The method comprises the steps of calculating a weight according to a training log: the weight of each user during model aggregation is calculated by using the training log of federated learning by a server, wherein in common federated learning, in the (t + 1) th epoch, the server aggregation model can be formalized as description, wherein theta t + 1 represents the global model of the (t + 1) th round, and theta t is the global model of the tth epoch. According to theinvention, contribution is calculated through the weight of a user during model aggregation, the calculation overhead is greatly reduced, the index-level overhead is reduced into linear overhead, model convergence can be accelerated, and the performance is improved; 2) the used training log is inherent information of federated learning, and extra privacy protection cost is avoided;.

Description

technical field [0001] The invention relates to efficient contribution evaluation in federated learning scenarios, which belongs to the field of machine learning and data analysis. Background technique [0002] Federated Learning (Federated Learning) is an emerging artificial intelligence basic technology. It was first proposed by Google in 2016. It was originally used to solve the problem of updating models locally for end users of Android phones. Its design goal is to guarantee big data exchange. Under the premise of information security, protection of terminal data and personal data privacy, and compliance with laws and regulations, efficient machine learning is carried out among multiple participants or computing nodes. [0003] A major feature of federated learning is that it better solves the data island problem, and after the training process, each user's contribution to the model can be recorded in the permanent data recording mechanism, and the actual effect will be...

AI Technical Summary

Problems solved by technology

Existing work usually uses the Shapley value to calculate the contribution of each training sample to the model training, but this cannot be directly used in the federated system. The main reason is that calculating the Shapley value requires exponential computing and communication overhead, which in reality Usually not possible in federated learning systems

Images