Personalisation in Federated Learning (FL) aims to adapt a global model to heterogeneous client distributions so that each client achieves better local performance without centralising private data.
Experimental setup
- Dataset: CIFAR-10.
- Model: ResNet-18.
- Topology: 10 clusters × 10 clients (100 clients total).
- Local skew: each client is highly imbalanced — 90% of samples belong to a single dominant class.
- Metrics:
- Personalised Accuracy — accuracy on a test split matching the client’s local distribution.
- Generalisation Accuracy — accuracy on a global test set disjoint from all clients.
Reproduce this table:
python scripts/pers_cifar10_script.py > personalization_log.txt &
The results illustrate the trade-off between Finetune’s strong adaptation to local samples at the cost of generalization, FedAMP’s balanced performance on both metrics, and FedAvg’s baseline non-personalized performance.
| Method | Personalised Accuracy | Generalisation Accuracy |
|---|---|---|
| FedAvg | 0.801 | 0.794 |
| Finetune | 0.905 | 0.703 |
| FedAMP | 0.840 | 0.782 |
Finetune achieved strong personalization but at the cost of reduced generalization. FedAMP, in contrast, outperformed FedAvg in local adaptation while preserving generalization. These insights highlight the core trade-off in federated learning personalization between model specialization and generalization. Our framework supports a structured investigation of this balance, offering tools for reproducible and systematic experimentation.