IBM/differential-privacy-library

Privacy-preserving parametrisation

Opened this issue · 0 comments

At present, when parameters such as bounds and data_norm are not specified by the user, the values are simply read from the data, a clear violation of differential privacy (and flagged as such with a PrivacyLeakWarning). It is possible, however, to estimate these parameters using histograms, thanks to their low-sensitivity.

With clip_to_norm and clip_to_bounds a feature from diffprivlib 0.3, implementing these estimations can be done with greater ease.

Similar implementations to this can be found here and here.