fedper

**FedPer** (Federated Personalization) is a **personalized federated learning approach that splits the model into shared base layers and private personalization layers** — the base layers are federated (shared across clients), while the top layers remain local to each client for personalized predictions. **How FedPer Works** - **Base Layers**: Lower/feature extraction layers are shared and aggregated globally via FedAvg. - **Personalization Layers**: Top layers (typically the classifier head) stay local — not shared. - **Training**: Each client trains the full model, sends only base layer updates, and keeps personalization layers private. - **Split Point**: Choose which layers to share vs. keep private based on the task and heterogeneity. **Why It Matters** - **Personalization**: Each client has a personalized model that fits their local data distribution. - **Shared Features**: Base layers learn general features from all clients' data — more robust feature extraction. - **Privacy**: Personalization layers are never communicated — additional privacy for local patterns. **FedPer** is **shared foundation, personal touch** — federating common feature learning while keeping task-specific decisions private and personalized.