ruby/tensorflow_privacy
1
0
Fork 0
forked from 626_privacy/tensorflow_privacy
Code Pull requests Activity

Add logging for secret sharer exposure computation.

Browse source 
PiperOrigin-RevId: 467771239
This commit is contained in:
Shuang Song 2022-08-15 15:06:18 -07:00 committed by A. Unique TensorFlower
parent 8a449aaa27
commit 40d73ed240
1 changed files with 19 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

24
tensorflow_privacy/privacy/privacy_tests/secret_sharer/exposures.py
View file

@ -13,7 +13,9 @@
# limitations under the License.
"""Measuring exposure for secret sharer attack."""
from typing import Iterable, TypeVar, Mapping
import logging
from typing import Mapping, Sequence, TypeVar
import numpy as np
from scipy import stats
@ -22,8 +24,8 @@ _KT = TypeVar('_KT')
def compute_exposure_interpolation(
perplexities: Mapping[_KT, Iterable[float]],
perplexities_reference: Iterable[float]) -> Mapping[_KT, Iterable[float]]:
perplexities: Mapping[_KT, Sequence[float]],
perplexities_reference: Sequence[float]) -> Mapping[_KT, Sequence[float]]:
"""Gets exposure using interpolation.
Args:
@ -37,6 +39,12 @@ def compute_exposure_interpolation(
The exposure of every secret measured using interpolation (not necessarily
in the same order as the input), keyed in the same way as perplexities.
"""
logging.info(
'Will compute exposure (with interpolation) for '
'splits %s with %s examples using %s references.',
str(perplexities.keys()), str([len(p) for p in perplexities.values()]),
len(perplexities_reference))
# Get the keys in some fixed order which will be used internally only
# further down.
keys = list(perplexities)
@ -67,8 +75,8 @@ def compute_exposure_interpolation(
def compute_exposure_extrapolation(
perplexities: Mapping[_KT, Iterable[float]],
perplexities_reference: Iterable[float]) -> Mapping[_KT, Iterable[float]]:
perplexities: Mapping[_KT, Sequence[float]],
perplexities_reference: Sequence[float]) -> Mapping[_KT, Sequence[float]]:
"""Gets exposure using extrapolation.
Args:
@ -82,6 +90,12 @@ def compute_exposure_extrapolation(
The exposure of every secret measured using extrapolation, keyed in the same
way as perplexities.
"""
logging.info(
'Will compute exposure (with extrapolation) for '
'splits %s with %s examples using %s references.',
str(perplexities.keys()), str([len(p) for p in perplexities.values()]),
len(perplexities_reference))
# Fit a skew normal distribution using the perplexities of the references
snormal_param = stats.skewnorm.fit(perplexities_reference)