Module dp 

Differential Privacy Module (MLOPS-015)

DP-SGD implementation following Abadi et al. (2016) for privacy-preserving training.

Toyota Way: Jidoka (Autonomation)

Built-in privacy protection stops data leakage. The privacy budget acts as an Andon system, halting training when privacy guarantees would be violated.

Example

use entrenar::optim::dp::{DpSgdConfig, DpSgd, PrivacyBudget};
use entrenar::optim::SGD;

let config = DpSgdConfig::new()
    .with_max_grad_norm(1.0)
    .with_noise_multiplier(1.1)
    .with_budget(PrivacyBudget::new(8.0, 1e-5));

let dp_sgd = DpSgd::new(SGD::new(0.01), config);

References

[3] Abadi et al. (2016) - Deep Learning with Differential Privacy [4] Mironov (2017) - Renyi Differential Privacy

Re-exports

pub use accountant::compute_rdp_gaussian;

pub use accountant::rdp_to_dp;

pub use accountant::RdpAccountant;

pub use budget::PrivacyBudget;

pub use config::DpSgdConfig;

pub use dp_sgd::DpSgd;

pub use error::DpError;

pub use error::Result;

pub use gradient::add_gaussian_noise;

pub use gradient::clip_gradient;

pub use gradient::grad_norm;

pub use utils::estimate_noise_multiplier;

pub use utils::privacy_cost_per_step;

Modules

accountant

RDP (Renyi Differential Privacy) accountant.

budget

Privacy budget tracking.

config

Configuration for DP-SGD.

dp_sgd

Differentially private SGD wrapper.

error

Error types for the Differential Privacy module.

gradient

Gradient operations for differential privacy.

utils

Utility functions for differential privacy.