Struct DifferentialPrivacyConfig 

pub struct DifferentialPrivacyConfig {

    pub target_epsilon: f64,
    pub target_delta: f64,
    pub noise_multiplier: f64,
    pub l2_norm_clip: f64,
    pub batch_size: usize,
    pub dataset_size: usize,
    pub max_steps: usize,
    pub noise_mechanism: NoiseMechanism,
    pub secure_aggregation: bool,
    pub adaptive_clipping: bool,
    pub adaptive_clip_init: f64,
    pub adaptive_clip_lr: f64,
    pub adaptive_clip_target_quantile: f64,
    pub adaptive_clip_noise_multiplier: f64,
    pub accounting_method: AccountingMethod,
    pub acknowledge_aggregate_clipping: bool,
}

Differential privacy configuration

Fields

target_epsilon: f64

Target privacy parameter epsilon (the budget that is enforced)

target_delta: f64

Delta at which epsilon is reported (typically << 1/n). This is a reporting parameter, not an additively consumed budget.

noise_multiplier: f64

Noise multiplier for gradient perturbation (sigma, relative to the clipping norm)

l2_norm_clip: f64

L2 norm clipping threshold applied to each per-example gradient

batch_size: usize

Expected batch size (used for the sampling probability)

dataset_size: usize

Dataset size for privacy accounting

max_steps: usize

Maximum number of training steps (enforced)

noise_mechanism: NoiseMechanism

Noise mechanism to use

secure_aggregation: bool

Enable secure aggregation (for federated learning). When set, the optimizer refuses to run unless a secure aggregation backend has been wired in, rather than silently training without it.

adaptive_clipping: bool

Enable adaptive clipping (Andrew et al. 2021, differentially private quantile estimation)

adaptive_clip_init: f64

Initial clipping threshold for adaptive clipping

adaptive_clip_lr: f64

Learning rate (geometric update rate) for adaptive clipping

adaptive_clip_target_quantile: f64

Target fraction of per-example gradients that should fall below the clipping threshold (Andrew et al. 2021 use 0.5)

adaptive_clip_noise_multiplier: f64

Noise multiplier applied to the privatized above-threshold count used by adaptive clipping. The count has sensitivity 1, so this is the standard deviation of the Gaussian added to it.

accounting_method: AccountingMethod

Privacy accounting method

acknowledge_aggregate_clipping: bool

Explicit acknowledgement that the caller understands the semantics of DifferentiallyPrivateOptimizer::dp_step, which clips an already aggregated gradient and therefore does not provide per-example differential privacy. Left false, that entry point returns an error.

Implementations

impl DifferentialPrivacyConfig

pub fn validate(&self) -> Result<()>

Validate the configuration.

Every parameter that can silently void a privacy guarantee is checked here: a zero noise multiplier, a non-positive clipping norm, a delta outside (0, 1), a batch larger than the dataset, and so on.

pub fn sampling_probability(&self) -> f64

Per-record sampling probability q = batch_size / dataset_size.

Trait Implementations

impl Clone for DifferentialPrivacyConfig

fn clone(&self) -> DifferentialPrivacyConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DifferentialPrivacyConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for DifferentialPrivacyConfig

fn default() -> Self

Returns the “default value” for a type.