FedBack: Controlling Participation in Federated Learning with Feedback

Feedback control of client participation in ADMM-based federated learning

Overview

Most federated learning algorithms pick a small random subset of clients in every training round. FedBack replaces this random selection with a deterministic, control-theoretic mechanism. Client participation is modelled as a discrete-time dynamical system, and an integral feedback controller adjusts each client’s participation threshold individually, based on how that client’s local optimization is evolving. A client only communicates when its local model has changed sufficiently, and the controller keeps each client’s long-run participation rate at a user-defined target.

This work was conducted by Michael Cummins, Güner Dilsad Er and Michael Mühlebach at the Max Planck Institute for Intelligent Systems.

Paper published at L4DC 2025 — Read in PMLR · arXiv


Method

FedBack block diagram: an integral controller sets each client's threshold delta, which drives event-triggered ADMM participation so that the participation rate L tracks the reference L_ref
FedBack closes a feedback loop around event-triggered ADMM: the controller sets each client's threshold so that its participation rate tracks the reference.

Key Features

  • Deterministic participation: clients take part when their local updates are informative, rather than being sampled at random
  • Integral feedback control: each client’s threshold is tuned online so its participation rate tracks a chosen target
  • Theoretical guarantees: the time-averaged participation rate converges to the target, every client keeps participating, and the algorithm converges globally
  • Efficiency: on federated image classification, FedBack improves communication and computational efficiency by up to 50% compared with random client selection

Results

Experiments on MNIST and CIFAR-10 with non-i.i.d. client data compare FedBack against FedAvg, FedProx and FedADMM, measuring the number of client participation events needed to reach a target accuracy. FedBack reaches the target with markedly fewer participation events than random-selection ADMM, and on CIFAR-10 at low participation rates it reaches accuracy targets that the averaging-based baselines do not.

Training curves for different participation rates:

Training curves of FedBack and baselines for different participation rates

Total number of client participation events required to reach target accuracy:

Table of client participation events needed by FedBack and baselines to reach target accuracy on MNIST and CIFAR-10

Code & Implementation

Open-source code to reproduce the paper’s experiments. See the README for setup and usage instructions.


Citation

@inproceedings{cummins2025feedback,
  title={Controlling Participation in Federated Learning with Feedback},
  author={Cummins, Michael and Er, Guner Dilsad and Muehlebach, Michael},
  booktitle={Proceedings of the 7th Annual Learning for Dynamics \& Control Conference},
  pages={174--186},
  year={2025},
  volume={283},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  url={https://proceedings.mlr.press/v283/cummins25a.html}
}