Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedMut: Generalized Federated Learning via Stochastic Mutation
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 FedMut, a federated learning method that dispatches stochastically mutated copies of the global model (mutated along the aggregated gradient direction with ±1 layer-wise signs) so the global model converges to a flat, well-generalizing region.
Problem
Federated learning under data heterogeneity (non-IID clients) suffers low global-model performance because FedAvg dispatches the same global model to all clients, which can get stuck in sharp minima that generalize poorly across clients.
Approach
Instead of dispatching the identical global model, the server generates K mutated models by adding or subtracting each layer's aggregated gradient (change from the previous global model) with random ±1 sign matrices constrained to average to the global model; the mutated models are trained locally so the aggregate converges to a flat region, while remaining compatible with secure aggregation and a
Key results
- FedMut reaches up to 11.02% accuracy improvement over baselines on CIFAR-10 with ResNet-18 at d=0.1 non-IID.
- On Shakespeare (LSTM), FedMut achieves 55.53% test accuracy versus 52.08% for FedAvg.
- Loss landscape visualization indicates the FedMut-trained ResNet-18 lies in a flatter region than the FedAvg-trained model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Mutation-based global training (FedMut) instead of FedAvg | CNN, ResNet-18, VGG-16 on CIFAR-10/CIFAR-100 · CIFAR-10/CIFAR-100 IID and non-IID (d=0.1,0.5,1.0) | Test accuracy (%): up to 11.02% gain on CIFAR-10 ResNet-18 d=0.1 | Helped |
| Mutation-based global training (FedMut) instead of FedAvg | LSTM · Shakespeare | Test accuracy (%): 55.53 vs 52.08 | Helped |
| Mutation-based global training vs. FedProx, FedGen, CluSamp baselines | CIFAR-10/CIFAR-100 across models · Table 1 settings | Test accuracy (%) | Helped |
| Increase mutation range alpha from 1 to 4 (ablation) | FedMut · CIFAR-10 ResNet-18 learning curves | Test accuracy (%) | Helped |
| Increase mutation range alpha to 5 (ablation) | FedMut · CIFAR-10 ResNet-18 learning curves | Test accuracy (%) | Hurt |
| Dynamic preference mutation with higher beta0 (early-stage gradient-direction bias) (ablation) | FedMut · CIFAR-10 learning curves | Convergence speed | Helped |
| Too large beta0 in dynamic preference mutation (ablation) | FedMut · CIFAR-10 learning curves | Training success | Hurt |
| Mutation-based training compatible with secure aggregation and no extra communication | FedMut · Discussion | Communication overhead | No clear effect |
Surprising findings
- Too large a mutation range (alpha=5) causes FedMut to fail to train a usable model.
- Clustered sampling (CluSamp) underperforms FedAvg on Shakespeare (49.74 vs 52.08).
Limitations
- Too large mutation range prevents convergence, so alpha must be tuned.
- Focuses only on global model inference performance; personalized and clustered FL not considered.
- Evaluated only on homogeneous model architectures.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Shakespeare
Benchmarks: CIFAR-10 CIFAR-100 Shakespeare
Metrics: test accuracy convergence curves loss landscape visualization
Baselines: FedAvg FedProx FedGen ClusteredSampling (CluSamp)
Builds on: FedAvg FedProx SCAFFOLD Dirichlet non-IID partitioning
Closest papers by meaning
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.87)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.86)
- Local Superior Soups: A Catalyst for Model Merging in Cross-Silo Federated Learning (NeurIPS 2024 · similarity 0.86)
- Improving Generalization in Federated Learning with Highly Heterogeneous Data via Momentum-Based Stochastic Controlled Weight Averaging (ICML 2025 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · similarity 0.86)
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.