Core Concepts
The proposed scheme combines offline data sharing and approximate gradient coding to mitigate the effects of label heterogeneity and client straggling in federated learning, while enabling a deliberate trade-off between privacy and utility.
Abstract
The content focuses on addressing the challenges of non-IID data and stragglers/dropouts in federated learning (FL). It introduces a privacy-flexible paradigm that models parts of the clients' local data as non-private, offering a more versatile and business-oriented perspective on privacy.
The key components of the proposed scheme are:
Offline data sharing: Clients share some of their non-private data with each other to reduce the statistical imbalances resulting from label heterogeneity and create redundancy in the training datasets.
Approximate gradient coding: This coding method is designed to provide an unbiased estimate of the central model update rule of gradient descent in the presence of stragglers, and the authors show that it reduces the variance of the obtained estimate, suggesting faster convergence.
The authors provide theoretical analysis and numerical simulations using the MNIST dataset to demonstrate that their approach enables achieving a deliberate trade-off between privacy and utility, leading to improved model convergence and accuracy while using an adaptable portion of non-private data.
Stats
The total number of training examples is M = 300, with K = 30 examples from each of the L = 10 classes.
The number of clients is N = 10.
The straggling probability is parameterized by p.