differential_privacy

`differential_privacy` ¶

Classes:

Name	Description
`DifferentialPrivacyHyperparams`	Hyperparameters for differential privacy during training.

`DifferentialPrivacyHyperparams` `pydantic-model` ¶

Bases: Parameters

Hyperparameters for differential privacy during training.

These parameters configure differential privacy (DP) training using DP-SGD algorithm. When enabled, they provide formal privacy guarantees by adding calibrated noise during training.

Fields:

dp_enabled (bool)
epsilon (float)
delta (AutoFloatParam)
per_sample_max_grad_norm (float)

`dp_enabled` `pydantic-field` ¶

Enable differentially-private training with DP-SGD.

`epsilon` `pydantic-field` ¶

Target privacy budget -- lower values provide stronger privacy. Must be > 0.

`delta` `pydantic-field` ¶

Probability of accidentally leaking information. Should be much smaller than 1/n where n is the number of training records. Setting to 'auto' uses delta of 1/n^1.2. Must be in [0, 1) or 'auto'.

`per_sample_max_grad_norm` `pydantic-field` ¶

Maximum L2 norm for per-sample gradient clipping. Must be > 0.

differential_privacy