Robust Federated Aggregation (RFA)

This is a TFF implmentation of the RFA algorithm, a federated learning algorithm robust to corruptions in client updates. As detailed here, the robust aggregation is achieved by computing an approximate Geometric median with the smoothed Weiszfeld algorithm. See the paper for details.

Use

This directory gives a function build_robust_federated_aggregation_process which builds a federated aggregation process implementing the RFA algorithm. It works as a drop-in replacement to tff.learning.build_federated_averaging_process for vanilla FedAvg. It can be used in the TFF EMIST tutorial as follows:

from federated_research.robust_aggregation import build_robust_federated_aggregation_process
# dataset setup, model setup, etc. from tutorial
....
# change definition of iterative process from
# iterative_process = tff.learning.build_federated_averaging_process(model_fn)
# to
iterative_process = build_robust_federated_aggregation_process(model_fn)
# Rest of the code remains unchanged

In addition, the directory gives a standalone function build_stateless_robust_aggregation build a robust aggregation oracle as an instance of tff.utils.StatefulAggregateFn.

Reproducing Experimental Results

To reproduce experimental results from the paper, see the official repository.

Citation

If you find this implementation useful, please cite the following paper:

@article{pillutla2019robust,
title={{R}obust {A}ggregation for {F}ederated {L}earning},
author={Pillutla, Krishna and  Kakade, Sham M. and Harchaoui, Zaid},
journal={Preprint},
year={2019}
}