tf_privacy.compute_dp_sgd_privacy

Compute epsilon based on the given hyperparameters.

tf_privacy.compute_dp_sgd_privacy(
    n, batch_size, noise_multiplier, epochs, delta
)

This function is deprecated. It does not account for doubling of sensitivity with microbatching, and assumes Poisson subsampling, which is rarely used in practice. (See "How to DP-fy ML: A Practical Guide to Machine Learning with Differential Privacy", https://arxiv.org/abs/2303.00654, Sec 5.6.) Most users should call compute_dp_sgd_privacy_statement, which provides appropriate context for the guarantee (see the reporting recommendations in "How to DP-fy ML", Sec 5.3). If you need a numeric epsilon value under specific assumptions, it is recommended to use the dp_accounting libraries directly to compute epsilon, with the precise and correct assumptions of your application.

n Number of examples in the training data.
batch_size Batch size used in training.
noise_multiplier Noise multiplier used in training.
epochs Number of epochs in training.
delta Value of delta for which to compute epsilon.

A 2-tuple containing the value of epsilon and the optimal RDP order.