StatMix: Data augmentation method that relies on image statistics in federated learning

TL;DR

StatMix is a novel data augmentation technique leveraging image statistics to enhance federated learning performance while maintaining data privacy, demonstrated through improved accuracy on CIFAR datasets.

Contribution

This paper introduces StatMix, a new image statistics-based augmentation method specifically designed for federated learning scenarios to improve model accuracy.

Findings

StatMix improves average accuracy in federated learning experiments.

StatMix shows some benefits even in non-federated setups.

Empirical validation on CIFAR-10 and CIFAR-100 datasets.

Abstract

Availability of large amount of annotated data is one of the pillars of deep learning success. Although numerous big datasets have been made available for research, this is often not the case in real life applications (e.g. companies are not able to share data due to GDPR or concerns related to intellectual property rights protection). Federated learning (FL) is a potential solution to this problem, as it enables training a global model on data scattered across multiple nodes, without sharing local data itself. However, even FL methods pose a threat to data privacy, if not handled properly. Therefore, we propose StatMix, an augmentation approach that uses image statistics, to improve results of FL scenario(s). StatMix is empirically tested on CIFAR-10 and CIFAR-100, using two neural network architectures. In all FL experiments, application of StatMix improves the average accuracy, compared to the baseline training (with no use of StatMix). Some improvement can also be observed in non-FL setups.