L-DAWA: Layer-wise Divergence Aware Weight Aggregation in Federated Self-Supervised Visual Representation Learning

TL;DR

This paper introduces L-DAWA, a novel federated learning aggregation method that mitigates client bias and divergence in self-supervised visual representation learning, leading to state-of-the-art results on multiple datasets.

Contribution

The paper proposes a layer-wise divergence-aware weight aggregation strategy for federated SSL, addressing client divergence issues during model aggregation.

Findings

Achieves new state-of-the-art performance on CIFAR-10/100 and Tiny ImageNet.

Effectively reduces client bias and divergence during federated aggregation.

Improves robustness and quality of visual representations in federated SSL.

Abstract

The ubiquity of camera-enabled devices has led to large amounts of unlabeled image data being produced at the edge. The integration of self-supervised learning (SSL) and federated learning (FL) into one coherent system can potentially offer data privacy guarantees while also advancing the quality and robustness of the learned visual representations without needing to move data around. However, client bias and divergence during FL aggregation caused by data heterogeneity limits the performance of learned visual representations on downstream tasks. In this paper, we propose a new aggregation strategy termed Layer-wise Divergence Aware Weight Aggregation (L-DAWA) to mitigate the influence of client bias and divergence during FL aggregation. The proposed method aggregates weights at the layer-level according to the measure of angular divergence between the clients' model and the global model. Extensive experiments with cross-silo and cross-device settings on CIFAR-10/100 and Tiny ImageNet datasets demonstrate that our methods are effective and obtain new SOTA performance on both contrastive and non-contrastive SSL approaches.