FedGraph: an Aggregation Method from Graph Perspective

TL;DR

FedGraph introduces an adaptive aggregation method for federated learning that considers data distribution, model topology, and training dynamics to improve performance on non-i.i.d datasets.

Contribution

It proposes FedGraph, a novel aggregation strategy that dynamically adjusts weights based on data size, model topology, and training conditions, addressing non-i.i.d data challenges.

Findings

Outperforms state-of-the-art by 2.76 Dice score on FeTS datasets.

Effectively handles non-i.i.d data in federated learning.

Demonstrates improved model correlation understanding through topology graphs.

Abstract

With the increasingly strengthened data privacy act and the difficult data centralization, Federated Learning (FL) has become an effective solution to collaboratively train the model while preserving each client's privacy. FedAvg is a standard aggregation algorithm that makes the proportion of dataset size of each client as aggregation weight. However, it can't deal with non-independent and identically distributed (non-i.i.d) data well because of its fixed aggregation weights and the neglect of data distribution. In this paper, we propose an aggregation strategy that can effectively deal with non-i.i.d dataset, namely FedGraph, which can adjust the aggregation weights adaptively according to the training condition of local models in whole training process. The FedGraph takes three factors into account from coarse to fine: the proportion of each local dataset size, the topology factor of model graphs, and the model weights. We calculate the gravitational force between local models by transforming the local models into topology graphs. The FedGraph can explore the internal correlation between local models better through the weighted combination of the proportion each local dataset, topology structure, and model weights. The proposed FedGraph has been applied to the MICCAI Federated Tumor Segmentation Challenge 2021 (FeTS) datasets, and the validation results show that our method surpasses the previous state-of-the-art by 2.76 mean Dice Similarity Score. The source code will be available at Github.