Research map / Federated learning and privacy / Communication-efficient FL

FedSPU: Personalized Federated Learning for Resource-Constrained Devices with Stochastic Parameter Update

AAAI 2025 · Federated learning and privacy › Communication-efficient FL · approach: Efficient local training in FL · 5 citations (OpenAlex)

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 FedSPU, a personalized federated learning method for resource-constrained devices that keeps the full model on each client, randomly freezes neurons during local training, and adds an early stopping scheme to cut training cost.

Problem

Federated dropout reduces computation and communication for resource-constrained IoT clients but introduces bias on non-iid data and erodes local personalization, so the paper targets personalized FL under device heterogeneity.

Approach

Instead of pruning neurons to build smaller sub-models, each client keeps the full architecture and randomly freezes a fraction of neurons during local SGD, so the frozen portion stays personalized when only active parameters are aggregated; early stopping on a combined train/test loss further reduces cost.

Key results

• FedSPU improves final test accuracy by 4.45% on average over the best dropout baseline (Hermes).
• Early stopping reduces FedSPU training rounds by 25%–71%, with accuracy loss of 0.11% (EMNIST), 3.4% (Google Speech) and 9.2% (CIFAR10).
• Memory footprint is reduced by an average of 54% versus FedSelect at p_k=0.2 relative to full-model training methods.

Measured effects reported in the paper

Surprising findings

• FjORD's accuracy collapses to near chance on EMNIST (7.66% without ES, 0.08% with ES), suggesting instability of global right-to-left pruning under non-iid data.

Limitations

• Accuracy drops noticeably with early stopping on CIFAR10 (9.2%).
• Convergence theory relies on assumptions (smoothness, bounded gradient divergence) and is stated for non-convex local objectives only at a high level.
• Experiments use small CNNs on three datasets with 100 simulated clients.

Experimental setup

Datasets: EMNIST CIFAR10 Google Speech

Benchmarks: EMNIST CIFAR10 Google Speech

Metrics: mean final test accuracy total training time (hours) total size of parameter transmission (GB) memory footprint (MB) rounds to early stop

Baselines: FjORD Hermes FedMP PruneFL FedSelect Random Dropout

Builds on: federated dropout FedMP FjORD FedSelect

Closest papers by meaning

• Fed-QSSL: A Framework for Personalized Federated Learning under Bitwidth and Data Heterogeneity (AAAI 2024 · similarity 0.88)
• Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.87)
• FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.87)
• FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.87)
• Flexible Sharpness-Aware Personalized Federated Learning (AAAI 2025 · similarity 0.87)
• pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.86)