Publications

Harmonia: Accurate Federated Learning with All-Inclusive Dataset

2024 IEEE 17th International Conference on Cloud Computing (CLOUD), Shenzhen, China, 2024, pp. 302-304, doi: 10.1109/CLOUD62652.2024.00041.

Wonmi Choi, Juyoung Ahn, Yeonho Yoo, Chuck Yoo, Gyeongsik Yang

Download PDF

Abstract

Federated learning (FL) is an appealing model training technique that utilizes heterogeneous datasets and user devices, ensuring user data privacy. Existing FL research proposed device selection schemes to balance the computing speeds of devices. However, we observe that these schemes compromise prediction accuracy by ~57.7%. To solve this problem, we present Harmonia that enhances prediction accuracy, while also balancing the diverse computing speeds of devices. Our evaluation shows that Harmonia improves prediction accuracy by ~1.7x over existing schemes.

Paper summary

Keep a participant’s data in federated training

Coordinating devices owned by the same participant can include more of that participant’s data in training.

Federated learning can leave slow devices behind, excluding the data they hold. Harmonia groups devices belonging to the same owner and uses a faster device within each group to train on the group’s combined dataset.

Select a training leader within each owner’s group

Devices profile a training iteration, and the fastest eligible device becomes the group leader. Data moves to that leader within the owner’s group; unchanged data does not need to be transferred again. Leader selection is refreshed periodically, and the server combines group updates using FedAvg.

The grouping changes where a participant’s data is processed without pooling data across different owners.

Each owner’s devices form a group with a selected training leader.View full-size figure
Each owner’s devices form a group with a selected training leader.

Test heterogeneous devices and data

The prototype uses three Raspberry Pis and two Jetson devices divided between two owners. Experiments vary the distribution and quantity of data across devices and compare Harmonia with TiFL and FedAvg.

In the quantity-heterogeneity experiment, converged accuracy is 1.71× TiFL’s and 1.61× FedAvg’s. The experiment evaluates the benefit of including group datasets through leader-based training.

Harmonia’s converged-accuracy ratios in the quantity-heterogeneity experiment.
BaselineHarmonia / baseline accuracy
TiFL1.71×
FedAvg1.61×

Key takeaways

  1. Slow devices can contribute data without doing all the training. Harmonia shifts computation to a faster device within the same owner’s group.
  2. Device heterogeneity becomes a leadership decision. Periodic profiling adapts which device trains while preserving the boundary between participant groups.