WHALE-FL: Wireless and Heterogeneity Aware Latency Efficient Federated Learning over Mobile Devices via Adaptive Subnetwork Scheduling

Huai-an Su, Jiaxiang Geng, Liang Li, Xiaoqi Qin, Yanzhao Hou, Hao Wang, Xin Fu, Miao Pan

TL;DR

WHALE-FL is a novel federated learning framework that dynamically adjusts the size of neural network "subnetworks" trained on mobile devices. By balancing real-time hardware constraints (like battery and signal) with the model's actual learning needs, it accelerates training by up to 2.1x without sacrificing final accuracy.

What Problem Does This Paper Address?

Federated Learning (FL) allows AI models to be trained across thousands of mobile devices without ever moving user data to a central server. However, the "real world" is messy. Mobile devices are incredibly heterogeneous: an iPhone 16 has vastly different computing power than a Raspberry Pi. Furthermore, wireless conditions fluctuate, and a device might suddenly become slow if the user starts playing a resource-heavy game.

Existing solutions like HeteroFL or Federated Dropout try to solve this by assigning smaller "subnetworks" to weaker devices. The problem? These assignments are static. They assume a device's capacity never changes and ignore the training progress itself. If a high-end device hits a dead zone or a background task spikes its CPU, it becomes a "straggler," holding up the entire global training round. WHALE-FL addresses this by making subnetwork scheduling adaptive and dynamic.

Methods

The core of WHALE-FL is a novel adaptive subnetwork selection utility function. Instead of a fixed assignment, each device calculates a utility score every round based on two factors:

  1. System Efficiency (SE): This captures the "here and now" of the hardware. It estimates the latency for both computation and wireless transmission. If a device is on a slow LTE connection or its CPU is busy, the SE score drops, signaling the need for a smaller subnetwork to stay within the time budget.
  2. Training Efficiency (TE): This captures the "learning needs" of the model. WHALE-FL uses Federated Fisher Information (FedFIM) to measure how much a change in model weights actually affects the output. During "Critical Learning Periods"—where the model is learning rapidly—the TE score rises, encouraging the use of larger subnetworks to capture complex features.

A hyperparameter ($\beta$) allows developers to trade off between these two. The resulting utility value is normalized and mapped to one of five discrete subnetwork width levels. The server then aggregates these varying-sized updates using a nested averaging technique, ensuring that every parameter is updated by the devices that trained it.

Datasets

The authors evaluated WHALE-FL across four diverse tasks to prove its versatility:

  • Image Classification: CNN on the MNIST dataset and ResNet18 on CIFAR-10.
  • Language Modeling: A Transformer model trained on WikiText2.
  • Human Activity Recognition (HAR): A CNN trained on the HAR dataset.

All experiments used a "non-IID" (Independent and Identically Distributed) data partition, which mimics real-world scenarios where different users have very different data patterns (e.g., one user takes mostly photos of cats, another of cars).

Key Insights and Ideas

  • Fisher Information as a Compass: One of the most elegant ideas is using Fisher Information to detect the "Critical Learning Period." As shown in the paper's evaluation, subnetwork sizes naturally increase when the model is at a stage where it benefits most from higher capacity, and shrink when the model is near convergence and "fine-tuning" is sufficient.
Figure 2: Fisher information and subnetwork size level changes over training time (CNN@MNIST). From left to right, the performance of the user-side models on MacBookPro 2018, NVIDIA Jetson TX2, and Raspberry Pi 4, as well as the global model’s performance, are shown.
  • Dynamic Straggler Mitigation: Unlike static methods, WHALE-FL doesn't just "deal" with stragglers; it prevents them. By shrinking the workload of a device that is currently resource-constrained, it ensures the entire fleet finishes the round on time.
  • The Efficiency Trade-off: The paper demonstrates that prioritizing only system efficiency leads to fast but inaccurate models, while prioritizing only training efficiency leads to accurate but painfully slow training. WHALE-FL finds the "sweet spot" between the two.

Key Results

WHALE-FL consistently outperformed both standard Federated Averaging (FedAvg) and static subnetwork methods:

  • Speedup: Compared to FedAvg, it achieved training speedups of 1.3x to 2.1x to reach target accuracy.
  • Vs. Static Methods: It outperformed HeteroFL by up to 1.74x in speed while maintaining or even improving final testing accuracy.
  • Robustness: The framework proved effective across different hardware tiers, from high-end MacBook Pros to low-end Raspberry Pis.
Figure 1: Performance comparison of different FL training approaches under various learning tasks. Figures from left to right are CNN@MNIST, ResNet18@CIFAR10, Transformer@WikiText2, and CNN@HAR with non-IID datasets.

Limitations and Future Work

While WHALE-FL is a significant step forward, the authors and the architecture suggest a few areas for further exploration:

  • Hyperparameter Sensitivity: The performance relies on tuning $\beta$ and the Fisher Information window size ($D$). While the paper provides a sensitivity analysis, automating this tuning would make the system more "plug-and-play."
  • Width-based Only: The current implementation focuses on width-wise subnetwork shrinkage (reducing the number of channels). Future work could integrate depth-based shrinkage (skipping layers) for even greater flexibility.
  • Communication Overhead: While subnetworks reduce the size of the payload, calculating and transmitting Fisher Information adds a small amount of overhead that could be further optimized.

Overall, WHALE-FL provides a robust blueprint for making federated learning viable on the diverse and unpredictable hardware that lives in our pockets.

Summary updated

Abstract

As a popular distributed learning paradigm, federated learning (FL) over mobile devices fosters numerous applications, while their practical deployment is hindered by participating devices' computing and communication heterogeneity. Some pioneering research efforts proposed to extract subnetworks from the global model, and assign as large a subnetwork as possible to the device for local training based on its full computing and communications capacity. Although such fixed size subnetwork assignment enables FL training over heterogeneous mobile devices, it is unaware of (i) the dynamic changes of devices' communication and computing conditions and (ii) FL training progress and its dynamic requirements of local training contributions, both of which may cause very long FL training delay. Motivated by those dynamics, in this paper, we develop a wireless and heterogeneity aware latency efficient FL (WHALE-FL) approach to accelerate FL training through adaptive subnetwork scheduling. Instead of sticking to the fixed size subnetwork, WHALE-FL introduces a novel subnetwork selection utility function to capture device and FL training dynamics, and guides the mobile device to adaptively select the subnetwork size for local training based on (a) its computing and communication capacity, (b) its dynamic computing and/or communication conditions, and (c) FL training status and its corresponding requirements for local training contributions. Our evaluation shows that, compared with peer designs, WHALE-FL effectively accelerates FL training without sacrificing learning accuracy.
BibTeX
@inproceedings{su2025whale,
  author = {Huai-an Su and Jiaxiang Geng and Liang Li and Xiaoqi Qin and Yanzhao Hou and Hao Wang and Xin Fu and Miao Pan},
  title = {{WHALE-FL: Wireless and Heterogeneity Aware Latency Efficient Federated Learning over Mobile Devices via Adaptive Subnetwork Scheduling}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i19.34272},
}
Figure