RdpAccountant

class RdpAccountant(config: PrivacyConfig, orders: tuple[float, ...] | None = None)

기반 클래스: object

Track cumulative privacy loss using Renyi differential privacy.

This class wraps Google’s dp-accounting package while exposing only Flower-owned configuration, event, result, and serialization types.

메소드

compose(event[, count])	Compose one or more identical sampled Gaussian releases.
get_delta(epsilon)	Return cumulative delta at the supplied epsilon.
get_epsilon(delta)	Return cumulative epsilon at the supplied delta.
get_privacy_spent([delta])	Return cumulative privacy expenditure at delta.
load_state_dict(state)	Restore state produced by state_dict().
state_dict()	Return JSON-serializable accountant state.
would_exceed(event[, max_epsilon, count])	Return whether composing an event would exceed an epsilon limit.

속성

config	Return the privacy configuration.
num_releases	Return the number of composed private releases.

compose(event: GaussianPrivacyEvent, count: int = 1) → None

Compose one or more identical sampled Gaussian releases.

property config: PrivacyConfig

Return the privacy configuration.

get_delta(epsilon: float) → float

Return cumulative delta at the supplied epsilon.

get_epsilon(delta: float) → float

Return cumulative epsilon at the supplied delta.

get_privacy_spent(delta: float | None = None) → PrivacySpent

Return cumulative privacy expenditure at delta.

load_state_dict(state: dict[str, object]) → None

Restore state produced by state_dict().

The restored state must use the same privacy configuration as this accountant. Backend numeric state is rebuilt from Flower’s event ledger.

property num_releases: int

Return the number of composed private releases.

state_dict() → dict[str, object]

Return JSON-serializable accountant state.

would_exceed(event: GaussianPrivacyEvent, max_epsilon: float | None = None, count: int = 1) → bool

Return whether composing an event would exceed an epsilon limit.

This method does not mutate committed accountant state.