Performance Measurement of Federated Learning on Imbalanced Data

Abstract

AI often suffers from getting imbalanced data distribution as unequal samples in classes which will increase the bias of machine learning algorithms. This research aimed to study effects of skew data distribution towards development of data rebalancing on Federated learning (FL) in the future. This research sets left skewed distribution, right skewed distribution and symmetric distribution on Modified National Institute of Standards and Technology database (MNIST) to operate on Convolutional neural network (CNN) in FL mechanism. Then, FL’s performance for working on these imbalanced distributions was tested. Results showed that in overview, the symmetric, left skewed, and right skewed distribution were not different in accuracy but theses imbalanced distributions were different in accuracy from the balanced distribution which has equal samples in all classes at significant level of.05. Standard deviation (SD) of data distribution was directly correlated with FL’s accuracy in high level.