Research map / Continual learning / Online and replay-based
MultiSFL: Towards Accurate Split Federated Learning via Multi-Model Aggregation and Knowledge Replay
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 MultiSFL, a split federated learning framework that maintains multiple branch models with master-model aggregation and an adaptive knowledge replay strategy that samples features from assistant clients to mitigate catastrophic forgetting.
Problem
Split federated learning suffers from low accuracy under non-IID data due to gradient divergence and catastrophic forgetting, since only partial clients train per round and standard fixes assume a whole model on each client.
Approach
Combines multi-model federated training (branch models with master-model aggregation) with a replay mechanism in which the main server requests feature samples of under-represented classes from idle assistant clients, weighted by an accumulated class-score vector, and adapts the sampling proportion using the federated gradient norm.
Key results
- MultiSFL outperforms SFL by 23.71% test accuracy on CIFAR-10 with MobileNetV2 at Dirichlet beta=0.1, and abstract reports up to 23.25% improvement overall.
- On CIFAR-10 with VGG-16 at beta=0.1, MultiSFL reaches 78.33% vs 65.86% for SFL and 66.26% for FedAvg.
- MultiSFL achieves best accuracy in all reported settings across four models and four datasets, with smaller IID-to-non-IID gaps than baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MultiSFL instead of SFL | SFL with MobileNetV2 · CIFAR-10, β = 0.1 | test accuracy: 23.71% | Helped |
| Add knowledge replay using assistant-client features (ablation) | MultiSFL · CIFAR-10, ResNet18, non-IID and IID settings | test accuracy | Helped |
| Add knowledge replay using assistant-client features (ablation) | MultiSFL · CIFAR-10, ResNet18, non-IID and IID settings | learning stability | Helped |
| Adaptively adjust the feature-sampling proportion instead of fixing it (ablation) | MultiSFL with knowledge replay · CIFAR-10, ResNet18, non-IID and IID settings | test accuracy | Helped |
| Adaptively adjust the feature-sampling proportion instead of fixing it (ablation) | MultiSFL with knowledge replay · CIFAR-10, ResNet18, non-IID and IID settings | training speed | Helped |
Limitations
- Split position must be identical across all models, so heterogeneous model architectures are not supported.
- Security and robustness (backdoor attacks, fairness) not addressed.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 FEMNIST TinyImageNet
Benchmarks: CIFAR-10 CIFAR-100 FEMNIST TinyImageNet
Metrics: test accuracy communication overhead
Baselines: FedAvg SFL FedMut FedNTD FedExP
Builds on: SplitFed FedAvg MMFL (multi-model federated learning) CriticalFL / critical learning periods GradMA
Closest papers by meaning
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.88)
- Towards Straggler-Resilient Split Federated Learning: An Unbalanced Update Approach (NeurIPS 2025 · similarity 0.86)
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.86)
- FSL-SAGE: Accelerating Federated Split Learning via Smashed Activation Gradient Estimation (ICML 2025 · similarity 0.85)
- GAS: Generative Activation-Aided Asynchronous Split Federated Learning (AAAI 2025 · similarity 0.85)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.84)
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.