DPFedAVG

Implementation of the DPFedAVG: Differential Privacy Federated Averaging [DPFedAVG2017] algorithm.

Note

This implementation does not exactly follow the original paper, but it is a simplified version that uses the Opacus library to provide differential privacy guarantees.

References

[DPFedAVG2017]

Robin C. Geyer, Tassilo Klein, Moin Nabi. Differentially Private Federated Learning: A Client Level Perspective In ArXiv (2017). URL: https://arxiv.org/abs/1712.07557

Classes included in fluke.algorithms.dpfedavg

DPFedAVG
DPFedAVGClient
DPFedAVGServer

Classes

class fluke.algorithms.dpfedavg.DPFedAVGClient(index: int, train_set: FastDataLoader, test_set: FastDataLoader, optimizer_cfg: OptimizerConfigurator, loss_fn: Module, local_epochs: int = 3, fine_tuning_epochs: int = 0, clipping: float = 0, noise_mul: float = 1.1, max_grad_norm: float = 1.0, **kwargs: dict[str, Any])

Bases: Client

class fluke.algorithms.dpfedavg.DPFedAVGServer(model: Module, test_set: FastDataLoader, clients: Iterable[Client], weighted: bool = False, lr: float = 1.0, **kwargs: dict[str, Any])

Bases: Server

class fluke.algorithms.dpfedavg.DPFedAVG(n_clients: int, data_splitter: DataSplitter, hyper_params: DDict | dict[str, Any], **kwargs: dict[str, Any])

Bases: CentralizedFL