Research map / Federated learning and privacy / Personalized and heterogeneous FL

DapperFL: Domain Adaptive Federated Learning with Model Fusion Pruning for Edge Devices

NeurIPS 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 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 DapperFL, a heterogeneous federated learning framework that prunes client models using fused local and global knowledge (MFP), adds a representation-norm regularizer (DAR), and aggregates heterogeneous pruned models, evaluated on Digits and Office Caltech.

Problem

Federated learning on heterogeneous edge devices suffers both from system heterogeneity (resource-constrained clients cannot train full models) and domain shifts across clients' local data, and existing methods address these only separately.

Approach

Prune each client's model with a mask computed from a fusion of the global model and one-epoch fine-tuned local model (dynamic fusion factor), regularize the pruned encoder's representation with a squared L2 norm, and aggregate by restoring pruned structure with the previous global model's weights before weighted averaging.

Key results

• DapperFL achieves 74.30% global accuracy on Digits (0.13% above runner-up FPL at 74.17%) and 67.75% on Office Caltech (2.28% above runner-up FedSR at 65.47%).
• Ablation on Office Caltech: removing both MFP and DAR drops accuracy to 62.65%; removing only DAR gives 64.88%; removing only MFP gives 66.28%.
• Pruning ratios of 0.2 to 0.8 reduce model volume, with personalized ratios giving reductions of 20% to 80% across client levels.

Measured effects reported in the paper

Surprising findings

• On Office Caltech, DapperFL at pruning ratio 0.4 (67.71%) is reported to exceed ratio 0.2 (66.80%), which the authors attribute to over-parameterization of ResNet18 for this task.

Limitations

• Introduces four hyperparameters (alpha_0, alpha_min, epsilon, gamma) that affect performance and are tuned manually or by Bayesian search.
• Assumes server knows each device's system capability and pruning ratios are set accordingly.
• Evaluated only on two small image benchmarks with a simulated 10-client setup and small-scale ResNet backbones.

Experimental setup

Datasets: Digits Office Caltech

Benchmarks: Digits Office Caltech

Metrics: top-1 accuracy (average across domains) number of parameters FLOPs

Baselines: FedAvg MOON FedSR FPL FedDrop FedProx FedMP NeFL

Builds on: FedAvg FedMP FedSR FPL

Closest papers by meaning

• FedLPS: Heterogeneous Federated Learning for Multiple Tasks with Local Parameter Sharing (AAAI 2024 · similarity 0.88)
• Cluster Based Heterogeneous Federated Foundation Model Adaptation and Fine-Tuning (AAAI 2025 · similarity 0.86)
• FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
• Overcoming Data and Model heterogeneities in Decentralized Federated Learning via Synthetic Anchors (ICML 2024 · similarity 0.86)
• Adaptive Latent-Space Constraints in Personalized Federated Learning (NeurIPS 2025 · similarity 0.86)
• Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.86)