DifferentialPrivacyClientSideAdaptiveClipping

class DifferentialPrivacyClientSideAdaptiveClipping(strategy: Strategy, noise_multiplier: float, num_sampled_clients: int, initial_clipping_norm: float = 0.1, target_clipped_quantile: float = 0.5, clip_norm_lr: float = 0.2, clipped_count_stddev: float | None = None)

기반 클래스: DifferentialPrivacyAdaptiveBase

Strategy wrapper for central DP with client-side adaptive clipping.

Use adaptiveclipping_mod modifier at the client side.

In comparison to DifferentialPrivacyServerSideAdaptiveClipping, which performs clipping on the server-side, DifferentialPrivacyClientSideAdaptiveClipping expects clipping to happen on the client-side, usually by using the built-in adaptiveclipping_mod.

매개변수:

• strategy (Strategy) – The strategy to which DP functionalities will be added by this wrapper.

• noise_multiplier (float) – The noise multiplier for the Gaussian mechanism for model updates.

• num_sampled_clients (int) – The number of clients that are sampled on each round.

• initial_clipping_norm (float) – The initial value of clipping norm. Defaults to 0.1. Andrew et al. recommends to set to 0.1.

• target_clipped_quantile (float) – The desired quantile of updates which should be clipped. Defaults to 0.5.

• clip_norm_lr (float) – The learning rate for the clipping norm adaptation. Defaults to 0.2. Andrew et al. recommends to set to 0.2.

• clipped_count_stddev (float) – The stddev of the noise added to the count of updates currently below the estimate. Andrew et al. recommends to set to expected_num_records/20

예제

Create a strategy:

strategy = fl.serverapp.FedAvg(...)

Wrap the strategy with the DifferentialPrivacyClientSideAdaptiveClipping wrapper:

dp_strategy = DifferentialPrivacyClientSideAdaptiveClipping(
    strategy, cfg.noise_multiplier, cfg.num_sampled_clients, ...
)

On the client, add the adaptiveclipping_mod to the client-side mods:

app = fl.client.ClientApp(mods=[adaptiveclipping_mod])

메소드

aggregate_evaluate(server_round, replies)	Aggregate MetricRecords in the received Messages.
aggregate_train(server_round, replies)	Aggregate ArrayRecords and MetricRecords in the received Messages.
configure_evaluate(server_round, arrays, ...)	Configure the next round of federated evaluation.
configure_train(server_round, arrays, ...)	Configure the next round of training.
start(grid, initial_arrays[, num_rounds, ...])	Execute the federated learning strategy.
summary()	Log summary configuration of the strategy.

aggregate_evaluate(server_round: int, replies: Iterable[Message]) → MetricRecord | None

Aggregate MetricRecords in the received Messages.

aggregate_train(server_round: int, replies: Iterable[Message]) → tuple[ArrayRecord | None, MetricRecord | None]

Aggregate ArrayRecords and MetricRecords in the received Messages.

configure_evaluate(server_round: int, arrays: ArrayRecord, config: ConfigRecord, grid: Grid) → Iterable[Message]

Configure the next round of federated evaluation.

configure_train(server_round: int, arrays: ArrayRecord, config: ConfigRecord, grid: Grid) → Iterable[Message]

Configure the next round of training.

start(grid: Grid, initial_arrays: ArrayRecord, num_rounds: int = 3, timeout: float = 3600, train_config: ConfigRecord | None = None, evaluate_config: ConfigRecord | None = None, evaluate_fn: Callable[[int, ArrayRecord], MetricRecord | None] | None = None, should_stop: Callable[[int, Result], bool] | None = None) → Result

Execute the federated learning strategy.

Runs the complete federated learning workflow for the specified number of rounds, including training, evaluation, and optional centralized evaluation.

매개변수:

• grid (Grid) – The Grid instance used to send/receive Messages from nodes executing a ClientApp.

• initial_arrays (ArrayRecord) – Initial model parameters (arrays) to be used for federated learning.

• num_rounds (int (default: 3)) – Maximum number of federated learning rounds to execute.

• timeout (float (default: 3600)) – Timeout in seconds for waiting for node responses. Outstanding responses from nodes that become unavailable are replaced with error replies.

• train_config (ConfigRecord, optional) – Configuration to be sent to nodes during training rounds. If unset, an empty ConfigRecord will be used.

• evaluate_config (ConfigRecord, optional) – Configuration to be sent to nodes during evaluation rounds. If unset, an empty ConfigRecord will be used.

• evaluate_fn (Callable[[int, ArrayRecord], Optional[MetricRecord]], optional) – Optional function for centralized evaluation of the global model. Takes server round number and array record, returns a MetricRecord or None. If provided, will be called before the first round and after each round. Defaults to None.

• should_stop (Callable[[int, Result], bool], optional) – Optional predicate to determine whether to stop training early. Called at the end of each round with the round number and accumulated results. Return True to stop and return the model and metrics collected so far. Not called for round zero. Results must not be modified. Defaults to None (no early stopping).

반환:

Results containing final model arrays and also training metrics, evaluation metrics and global evaluation metrics (if provided) from all rounds.

반환 형식:

Results

summary() → None

Log summary configuration of the strategy.