Struct DifferentialPrivacy 

pub struct DifferentialPrivacy {
    pub epsilon: f32,
    pub sensitivity: f32,
    pub noise_scale: f32,
}

Differential privacy for federated learning

Fields

epsilon: f32

Privacy budget (epsilon)

sensitivity: f32

Sensitivity

noise_scale: f32

Noise scale

Implementations

impl DifferentialPrivacy

pub fn new(epsilon: f32, sensitivity: f32) -> Self

Create a new differential privacy mechanism

pub fn add_noise(&self, tensor: &Tensor) -> Tensor

Add Gaussian noise to a tensor

pub fn clip_gradients( &self, gradients: &HashMap<String, Tensor>, max_norm: f32, ) -> HashMap<String, Tensor>

Clip gradients to bound sensitivity

pub fn privatize( &self, updates: &HashMap<String, Tensor>, ) -> HashMap<String, Tensor>

Apply differential privacy to aggregated updates