LoRA-federated

Each client trains LoRA adapter locally. Central aggregates adapters (FedAvg). Only adapter weights leave client.

Advertisement

Secure aggregation

Cryptographic aggregation: server sees only sum, not individual contributions. Meta's Federated Learning Framework.

Advertisement

DP layer

Add DP noise to per-client updates. Provable privacy guarantee.

Communication cost

LoRA adapters small (MB range). Feasible over WAN. Full fine-tune too expensive to federate.