This repository contains the code and experiments for the manuscript:

Ditto: Fair and Robust Federated Learning Through Personalization

Fairness and robustness are two important concerns for federated learning systems. In this work, we identify that robustness to data and model poisoning attacks and fairness, measured as the uniformity of performance across devices, are competing constraints in statistically heterogeneous networks. To address these constraints, we propose employing a simple, general framework for personalized federated learning, Ditto, and develop a scalable solver for it. Theoretically, we analyze the ability of Ditto to achieve fairness and robustness simultaneously on a class of linear problems. Empirically, across a suite of federated datasets, we show that Ditto not only achieves competitive performance relative to recent personalization methods, but also enables more accurate, robust, and fair models relative to state-of-the-art fair or robust baselines.

For each dataset, we provide links to downloadable datasets used in our experiments. We describe in our paper and the REAME files in separate ditto/data/$dataset folders on how these datasets are generated, and provide instructions and scripts on preprocessing and/or sampling data.

pip3 install -r requirements.txt

We provide a jupyter notebook that simulates the federated point estimation problem. To run that, make sure you are under the ditto folder, and

jupyter notebook

then open point_estimation.ipynb, and directly run the notebook cell by cell to reproduce the results.

(A subset of) Options in run.sh:

• dataset chosen from [femnist, fmnist, celeba, vehicle], where fmnist is short for Fashion MNIST.
• model should be the corresponding model of that dataset. You can find it the model name under flearn/models/$dataset/$model.py, and take $model.
• $optimizer chosen from ['ditto', 'apfl', 'ewc', 'kl', 'l2sgd', 'mapper', 'meta', 'fedavg', 'finetuning']
• fedavg is training global models, ditto with lam=0 corresponds to training separate local models
• $lambda is the lambda we use for ditto (can use dynamic lambdas by setting --dynamic_lam to 1)
• num_corrupted is the number of corrupted devices (see the total number of devices in paper)
• random_updates indicates whether we launch Attack 2 (Def 1 in paper)
• boosting indicates whether we launch Attack 3 (Def 1 in paper)
• If both random_updates and boosting is set to 0, then we default to Attack 1 (Def 1 paper)
• By default, we disable all robust baselines. If you want to test any of them, set --optimizer=fedavg, and set any of the robust baselines to 1 (chosen from gradient_clipping, krum, mkrum, median, k_norm, k_loss, fedmgda in run.sh). For fedmgda, one needs to set an additional fedmgda_eps hyperparameter, chosen from the continuous range of [0, 1]. For our experiments, we pick the best fedmgda_eps among {0, 0.1, 0.5, 1} based on validation performance on benign devices.

• Download datasets (link and instructions under ditto/data/fmnist/README.md)
• Fashion MNIST, Ditto, without attacks, lambda=1: bash run_fashion_clean_ditto_lam1.sh
• Fashion MNIST, Ditto, A1 (50% adversaries), lambda=1: bash run_fashion_a1_50_ditto_lam1.sh

• Download datasets (link and instructions under ditto/data/vehicle/README.md)
• Vehicle, Ditto, without attacks, lambda=1: bash run_vehicle_clean_ditto_lam1.sh
• Vehicle, Ditto, A1 (50% adversaries), lambda=1: bash run_vehicle_a1_50_ditto_lam1.sh