A buffered semi-asynchronous aggregation with age-variance staleness weighting ensures stable optimization. Experiments show FTTE reaches target accuracy faster, reducing training memory by 80% and communication by 69%.
Tackling Stragglers and Memory Limits in Edge-Based FL
FL enables collaborative training without raw data sharing, but its deployment on edge devices (such as phones, wearables, Internet of Things (IoT) sensors) faces two key challenges. First, stragglers, slow devices due to limited compute or connectivity, cause synchronization delays in synchronous FL or training instability in asynchronous FL due to stale updates. Second, strict memory and communication limits make full-model training infeasible for many devices.
Prior work on compression or staleness heuristics often degrades accuracy or requires careful tuning. This paper introduced FTTE, a semi-asynchronous framework that enforces global memory constraints via sparse, server-selected updates and uses age-variance staleness weighting to ensure stable, resource-efficient FL on heterogeneous edge devices.
Enabling FL on Memory-Limited Edge Devices
Existing FL approaches fail to jointly address the three core challenges of edge-dominated networks: severe device memory limitations, high communication costs, and training delays caused by stragglers. FTTE overcomes these gaps through a unified design that combines intelligent parameter selection with a sparse semi-asynchronous aggregation mechanism, creating a practical solution for real-world edge intelligence.
To guarantee that even the weakest device can participate, FTTE first enforces a global memory budget equal to the smallest memory capacity across all clients. The server then analyzes the full model to estimate which individual parameters, if updated, would contribute most to accuracy. It selects a sparse subset of trainable parameters to maximize the estimated accuracy while staying strictly within the global memory limit.
All other parameters are frozen. Consequently, every client stores, trains, and transmits only this compact subset. This single design choice simultaneously reduces on-device memory usage by up to 80% and communication payload by up to 69%, as freezing parameters eliminates their transfer entirely.
Rather than waiting for all clients to finish or updating after every single response, FTTE maintains a fixed-capacity buffer on the server. Available clients receive the sparse parameter subset, perform local training, and send back only their updates to the buffer. Only when the buffer reaches capacity does the server aggregate the updates and broadcast the improved model. This buffered approach decouples training progress from individual stragglers, preventing slow devices from blocking the entire system.
A key innovation is how FTTE weights each buffered update before aggregation. Each update receives a weight that decreases with both its age (the number of rounds since it was computed) and its statistical variance (how much it deviates from the current global model). Updates that are both old and highly divergent are effectively suppressed. This dual mechanism provides robust stability against stragglers and non-identical data distributions, enabling faster convergence than prior methods while scaling to over 500 clients.
Convergence, Efficiency, and Scalability Results
The authors evaluate FTTE through extensive simulations and real hardware experiments. They use pretrained models optimized for a 64 megabytes (MB) memory budget on TinyImageNet, with downstream datasets including Canadian Institute for Advanced Research (CIFAR)-10, Oxford Pets, Oxford Flowers, and Skin Cancer diagnosis. Data heterogeneity is controlled using a Dirichlet distribution. Experiments simulate 100 clients with moderate heterogeneity and 50 percent stragglers experiencing delays up to 30 s.
FTTE demonstrates significantly faster convergence than synchronous FL (SyncFL), asynchronous FL (AsyncFL), and FedBuff across all tested scenarios. While baseline methods often fail to reach target accuracy or oscillate, FTTE consistently achieves targets in fewer communication rounds.
In terms of memory and payload efficiency, FTTE reduces on-device training memory by 80 percent and communication payload by 69 percent compared to full-model update methods, while maintaining accuracy. Limiting updates to the last layer yields slightly lower resource usage but incurs up to a 7% drop in accuracy.
Under straggler conditions with up to 90 percent slow devices or delays up to 120 seconds, FTTE shows only modest increases in communication steps while SyncFL degrades sharply. The framework scales efficiently to 500 clients, achieving approximately 7.5 times faster convergence than SyncFL.
The age-variance staleness function outperforms age-only schemes by about 20%. Real-world tests on four Raspberry Pi devices confirm FTTE works effectively on modest edge hardware, attaining target accuracy within 13 to 19 communication rounds.
What This Means for Edge FL
FTTE presents a practical solution for FL on resource-constrained edge devices. By integrating memory-aware parameter selection, sparse updates, and age-variance staleness weighting, the framework enables robust semi-asynchronous training under extreme heterogeneity and straggler conditions. Extensive experiments demonstrate that FTTE achieves 81 percent faster convergence, reduces on-device memory by 80 percent, and cuts communication payload by 69 percent compared to existing methods.
It scales efficiently to 500 clients and handles up to 90% straggler cases while maintaining accuracy. Real-world tests on Raspberry Pi devices confirm its deployability. FTTE advances the state-of-the-art by jointly addressing robustness and resource efficiency in heterogeneous edge networks.
Journal Reference
Zewe, A. (2026, April). Enabling privacy-preserving AI training on everyday devices. MIT News | Massachusetts Institute of Technology. https://news.mit.edu/2026/enabling-privacy-preserving-ai-training-everyday-devices-0429
Tenison, I., Murphy, A., Beauville, C., & Kagal, L. (2025). FTTE: Enabling Federated and Resource-Constrained Deep Edge Intelligence. ArXiv.org. DOI:10.48550/arXiv.2510.03165, https://arxiv.org/abs/2510.03165
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.