FedACT: Concurrent Federated Intelligence across Heterogeneous Data Sources
A new algorithm optimizes concurrent model training across millions of heterogeneous devices.
Federated learning (FL) allows multiple devices to collaboratively train machine learning models without sharing raw data, but real-world deployments often require training many models simultaneously on the same pool of devices—leading to severe resource contention and inefficiency. Existing single-task FL optimizers fail in these multi-FL environments, especially when devices vary in compute, memory, and network capacity.
To solve this, researchers from the University of Louisiana at Lafayette, the University of Illinois Urbana-Champaign, and the University of North Texas propose FedACT, a novel scheduling algorithm that dynamically assigns heterogeneous devices to concurrent FL jobs. FedACT uses an alignment scoring mechanism to evaluate compatibility between each device's available resources (e.g., CPU, bandwidth) and the resource demands of each job, then prioritizes devices with higher scores while ensuring balanced participation across tasks—avoiding starvation or overuse. In extensive tests with diverse FL jobs and benchmark datasets, FedACT achieved up to 8.3× faster average job completion time and up to 44.5% higher model accuracy compared to prior methods. The work is published on arXiv (arXiv:2605.00011) and targets the growing need for efficient, multi-model federated intelligence in edge computing and IoT deployments.
- FedACT reduces average job completion time by up to 8.3× over state-of-the-art baselines
- Uses alignment scoring to match device resources (compute, memory, bandwidth) to job demands
- Improves model accuracy by up to 44.5% through balanced device participation across concurrent FL tasks
Why It Matters
Enables efficient, privacy-preserving multi-model training on resource-constrained edge devices, accelerating real-world FL deployment.