jax_privacy.matrix_factorization.banded.minsep_sensitivity_squared

jax_privacy.matrix_factorization.banded.minsep_sensitivity_squared(strategy, min_sep, max_participations=None, n=None, skip_checks=False)[source]

Returns the sensitivity of the ColumnNormalizedBanded strategy.

With max_participations = 1 (and any min_sep, say min_sep = 1), this is the same as single participation.

Parameters:

  • strategy (ColumnNormalizedBanded) – The strategy matrix defining the mechanism.

  • min_sep (int) – The minimum separation between two participation of a worst-case client/sample. Note that we use the definition in [(Amplified) Banded Matrix Factorization: A unified approach to private training](https://arxiv.org/abs/2306.08153). For a user participating on iteration $i$ and then again on iteration $j$, the separation is $j -i$; that is, a min_sep of 1 allows participation on every iteration.

  • max_participations (int | None) – The maximum participation of a worst-case user. The default value None allows the max number of possible participations.

  • n (int | None) – Optional, the size of the matrix C (see coef above). If None, the size of the matrix is equal to the number of coefficients.

  • skip_checks (bool) – If True, don’t perform input verification which may not be supported in jitted contexts.

Return type:

int

Returns:

The sensitivity squared.