Research map / Federated learning and privacy / Communication-efficient FL
WHALE-FL: Wireless and Heterogeneity Aware Latency Efficient Federated Learning over Mobile Devices via Adaptive Subnetwork Scheduling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WHALE-FL, which lets each mobile device adaptively pick its subnetwork width each round using a utility combining Fisher-information-based training efficiency and system (compute+communication) efficiency, reducing FL training latency over heterogeneous devices.
Problem
Fixed-size subnetwork assignment in federated learning over heterogeneous mobile devices ignores time-varying wireless and computing conditions and changing FL training-stage requirements, causing long training delay.
Approach
A subnetwork selection utility multiplying a window-averaged local Fisher information (training efficiency) by a power of the round-time-based system efficiency, normalized and mapped to discrete width levels, so each device adapts its subnetwork size per round rather than always using its maximum feasible size.
Key results
- WHALE-FL reaches target accuracy 1.74x faster than HeteroFL on CNN@MNIST and 1.25x on ResNet-18@CIFAR-10 (1.21x on Transformer@WikiText-2, 1.06x on CNN@HAR).
- Versus FedAvg, WHALE-FL speeds up training by about 1.5x, 1.9x, 1.3x and 2.1x across the four tasks.
- Applying WHALE to FedRolex gives 1.75x speedup on CNN@MNIST and improves final accuracy on CIFAR-10 from 72.52% to 79.57%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · CNN@MNIST | training speedup: 1.74x | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · ResNet@CIFAR10 | training speedup: 1.25x | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · Transformer@WikiText2 | training speedup: 1.21x | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · CNN@HAR | training speedup: 1.06x | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · CNN@MNIST | final accuracy: 87.42% -> 89.29% | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · ResNet@CIFAR10 | final accuracy: 71.65% -> 75.32% | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · Transformer@WikiText2 | final accuracy: 37.40% -> 39.28% | Helped |
| replace fixed HeteroFL subnetwork assignment with adaptive subnetwork scheduling | HeteroFL · CNN@HAR | final accuracy: 88.86% -> 91.38% | Helped |
| use training-efficiency utility only instead of the joint utility (ablation) | WHALE-FL · ResNet18@CIFAR10 | time to target accuracy | Hurt |
| use system-efficiency utility only instead of the joint utility (ablation) | WHALE-FL · ResNet18@CIFAR10 | testing accuracy | Hurt |
| set β to 1 rather than the balanced value β=2 (ablation) | WHALE-FL · CNN@MNIST and Transformer@WikiText2 | testing accuracy: β = 1 | Helped |
| set β to 5 rather than the balanced value β=2 (ablation) | WHALE-FL · CNN@MNIST and Transformer@WikiText2 | testing accuracy: β = 5 | Hurt |
Limitations
- Relies on a developer-specified hyperparameter beta, window size D, and threshold U_th that need tuning.
- Experiments use a small testbed of 20 devices with simulated wireless rates (LTE, Bluetooth, Wi-Fi 5 fixed rates).
Experimental setup
Datasets: MNIST CIFAR-10 HAR WikiText-2
Benchmarks: MNIST CIFAR-10 WikiText-2 HAR
Metrics: training latency speedup to target accuracy final testing accuracy
Baselines: FedAvg HeteroFL FedDropout FedRolex
Builds on: HeteroFL Federated Dropout FedRolex FedAvg FedFIM
Closest papers by meaning
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.88)
- SparsyFed: Sparse Adaptive Federated Learning (ICLR 2025 · similarity 0.87)
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.87)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.87)
- FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · similarity 0.87)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.