ruby/tensorflow_privacy
1
0
Fork 0
forked from 626_privacy/tensorflow_privacy
Code Pull requests Activity
789 commits 2 branches 0 tags 3.1 MiB
Commit graph

464 commits

Author SHA1 Message Date
Shuang Song
08364adcb7 Allow squared loss to take in labels and predictions of the same number of elements but different shapes. 
PiperOrigin-RevId: 474059427
 2022-09-13 10:32:58 -07:00
Yilei Yang
ebae6c086e Make this code compatible with Python 3.10. 
PiperOrigin-RevId: 473313795
 2022-09-09 12:20:05 -07:00
Chen Qian
715fd1a670 Code changes to get ready for an incoming Keras optimizer migration. 
Because the code subclasses the legacy Keras optimizer, we should explicitly use the legacy optimizer.

PiperOrigin-RevId: 473092233
 2022-09-08 14:56:56 -07:00
Steve Chien
407e5c8e11 Clarify logic in Keras version of DP-SGD optimizer, and add a unit test involving clipping on multiple variables. 
PiperOrigin-RevId: 472559697
 2022-09-06 14:36:43 -07:00
Steve Chien
ed73077b60 Change version to 0.8.5. (Previously incorrectly skipped ahead to 0.8.6) 
PiperOrigin-RevId: 471118348
 2022-08-30 16:28:14 -07:00
Steve Chien
875b7f46bd Automated rollback of commit cff47686f6 
PiperOrigin-RevId: 471104040
 2022-08-30 15:23:08 -07:00
A. Unique TensorFlower
cff47686f6 Changes DPOptimizerClass to generically accept and use any dp_sum_query. 
This enables creation of generic DPOptimizers by user's passing queries. The most common Gaussian query is automatically performed for convenience and backwards compatibility.

Byproducts of this update:
-ensures consistent implementations between the internal (and legacy) `get_gradients` and newer `_compute_gradients` for all queries.
-refactors for python readability.

PiperOrigin-RevId: 470883774
 2022-08-29 20:22:40 -07:00
Steve Chien
ed16033a92 Update pinned commit of dp-accounting library, update dependency versions, and increase version to 0.8.6. 
PiperOrigin-RevId: 470334560
 2022-08-26 14:30:16 -07:00
Shuang Song
9f4feade7d Add more documentation for gradient_accumulation_steps in keras optimizer. 
PiperOrigin-RevId: 469310667
 2022-08-22 16:16:46 -07:00
Steve Chien
9e25eee68b Update remaining DPQuery tests to TF2. 
PiperOrigin-RevId: 468793518
 2022-08-19 15:08:15 -07:00
Steve Chien
fd64be5b5b Update several DPQuery tests to TF v2. 
PiperOrigin-RevId: 468763153
 2022-08-19 12:40:49 -07:00
Steve Chien
d6ad59226d Update tests for optimizer classes to TF 2. 
PiperOrigin-RevId: 468587323
 2022-08-18 17:38:01 -07:00
Steve Chien
5dd11fcdd6 Add import of log_loss in keras_evaluation.py. 
PiperOrigin-RevId: 468294581
 2022-08-17 14:31:20 -07:00
Michael Reneer
052f9a3128 Update the version of numpy to 1.23.2. 
* Updated the numpy version.
* Synced the pandas version.

In Python 3.10, if you invoke `pip install pandas~=1.1.4 numpy~=1.21.4` and then `import pandas` you get the following error:

```
>>> import pandas
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/tmp/venv/lib/python3.10/site-packages/pandas/__init__.py", line 30, in <module>
    from pandas._libs import hashtable as _hashtable, lib as _lib, tslib as _tslib
  File "/tmp/venv/lib/python3.10/site-packages/pandas/_libs/__init__.py", line 13, in <module>
    from pandas._libs.interval import Interval
  File "pandas/_libs/interval.pyx", line 1, in init pandas._libs.interval
ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject
```

I believe that this is the cause of the issue https://github.com/scikit-learn-contrib/hdbscan/issues/457#issuecomment-773671043

PiperOrigin-RevId: 467952859
 2022-08-16 10:02:07 -07:00
Shuang Song
40d73ed240 Add logging for secret sharer exposure computation. 
PiperOrigin-RevId: 467771239
 2022-08-15 15:06:42 -07:00
Galen Andrew
8a449aaa27 Correct discrepancy between tensorflow-probability versions in requirements.txt vs setup.py. 
PiperOrigin-RevId: 467326193
 2022-08-12 17:04:55 -07:00
Galen Andrew
5a9866726d Change requirements for tensorflow-probability and pandas. 
PiperOrigin-RevId: 467220343
 2022-08-12 08:58:14 -07:00
Galen Andrew
ca077a5b12 Use calibrate_dp_mechanism from differential_privacy library instead of custom binary search. 
PiperOrigin-RevId: 466798182
 2022-08-10 15:19:44 -07:00
Shuang Song
a9abfbc244 Allow specifying loss function with string. 
PiperOrigin-RevId: 465333272
 2022-08-04 09:31:28 -07:00
Steve Chien
a8a5206841 Update TFP to version 0.8.2. 
PiperOrigin-RevId: 463664333
 2022-07-27 13:29:15 -07:00
Steve Chien
848cfc74c1 Add logistic regression functions to API. 
PiperOrigin-RevId: 463645193
 2022-07-27 12:05:06 -07:00
Shuang Song
17cd0c52bc Refactor: move loss computation utilities under privacy_tests. 
PiperOrigin-RevId: 463391913
 2022-07-26 11:49:40 -07:00
Michael Reneer
d16f020329 Fix usage of logging API. 
PiperOrigin-RevId: 463123944
 2022-07-25 10:48:31 -07:00
Galen Andrew
4cb0a11c4b Automated rollback of commit db292fc5d8 
PiperOrigin-RevId: 462171425
 2022-07-20 10:16:48 -07:00
Steve Chien
38fe4aa984 Changes to prepare for release of v0.8.1. 
Update WORKSPACE, setup.py, and requirements.txt to latest dp-accounting library release.

Update scipy version in setup.py.

Update version to 0.8.1.

PiperOrigin-RevId: 461944491
 2022-07-19 12:22:07 -07:00
A. Unique TensorFlower
2b5d5b6ef5 Add Positive Predictive Value as a metric for membership attack models performance on imbalanced data. 
PiperOrigin-RevId: 461390184
 2022-07-16 16:30:53 -07:00
Shuang Song
328795aa36 Add assert that the training is private for TF1 vectorized optimizer. 
In Keras training in TF 2.0+, compute_gradients() is not called but apply_gradients() is called. W/o calling compute_gradients() dp gradient is not computed, and a normal gradient is used.

PiperOrigin-RevId: 461021412
 2022-07-14 12:15:00 -07:00
A. Unique TensorFlower
e826ec717a Switch from a git_repository rule to http_archive for the DP accounting Bazel dependency. This is preferred, per https://docs.bazel.build/versions/main/external.html#repository-rules, to avoid depending on the system git (the HTTP downloader is build into Bazel). 
Also use the strip_prefix option to only pull in the accounting WORKSPACE, not the top-level Google DP project WORKSPACE. This allows us to align the import statements to work both when pulling in the `dp_acounting` dependency via Bazel and pip.

PiperOrigin-RevId: 459807060
 2022-07-08 12:07:17 -07:00
A. Unique TensorFlower
eaa7354f00 Internal change. 
PiperOrigin-RevId: 459620579
 2022-07-07 15:48:25 -07:00
Michael Reneer
94beccf10e Automated rollback of commit e32766cc73 
PiperOrigin-RevId: 458498524
 2022-07-01 10:34:33 -07:00
A. Unique TensorFlower
e32766cc73 Automated rollback of commit b16a0abf1c 
PiperOrigin-RevId: 458478847
 2022-07-01 08:52:40 -07:00
Michael Reneer
b16a0abf1c Update TFPs dp_accounting dependency. 
PiperOrigin-RevId: 458307242
 2022-06-30 13:57:42 -07:00
A. Unique TensorFlower
c665281c55 Implementation of Differentially Private Logistic Regression. 
PiperOrigin-RevId: 458266079
 2022-06-30 11:01:02 -07:00
Galen Andrew
77d962e0fc Use DP library's calibrate_dp_mechanism function to compute noise multiplier. 
calibrate_dp_mechanism is in dp_accounting's top level symbols, so after this change we can replace all of the dependencies on DP accounting subpackages with dp_accounting.

Tested by verifying that the resulting noise multiplier is the same before and after change when running all tests in multinomial_logistic_test.

PiperOrigin-RevId: 458235745
 2022-06-30 08:47:54 -07:00
Steve Chien
7a5f4eddb8 Internal change. 
PiperOrigin-RevId: 455173202
 2022-06-15 11:05:05 -07:00
Galen Andrew
db292fc5d8 Automated rollback of commit bb1b7a6a46 
PiperOrigin-RevId: 454995196
 2022-06-14 17:27:35 -07:00
Galen Andrew
bb1b7a6a46 Remove deprecated rdp_accountant. Please use RdpAccountant from the differential_privacy library. 
PiperOrigin-RevId: 454964256
 2022-06-14 15:02:34 -07:00
Galen Andrew
c4fea75675 Remove differential_privacy dp_event classes from TFP top-level symbols. 
PiperOrigin-RevId: 454238513
 2022-06-10 14:09:46 -07:00
Galen Andrew
125f82707a Migrate dependency of tree_aggregation_accountant_test on rdp_accountant to differential_privacy. 
PiperOrigin-RevId: 453989532
 2022-06-09 12:44:20 -07:00
Matthew Jagielski
6c0cc858e0 Add word2vec attack example to attack library. 
PiperOrigin-RevId: 453489150
 2022-06-07 11:46:55 -07:00
A. Unique TensorFlower
7d522e9273 Remove an indentation in a comment. 
PiperOrigin-RevId: 451008704
 2022-05-25 14:01:16 -07:00
A. Unique TensorFlower
5461f911a6 Bugfixes: 
1. Child classes of 'TrainedAttacker' now have a 'backend' parameter, so require an __init__() method.

PiperOrigin-RevId: 451005298
 2022-05-25 13:46:30 -07:00
A. Unique TensorFlower
95e527acfb Support models with unconnected layers and gradients when training using a DP vectorized optimizer. 
PiperOrigin-RevId: 450659644
 2022-05-24 05:36:58 -07:00
Steve Chien
f739f45299 Add .bazelversion file. 
PiperOrigin-RevId: 448082134
 2022-05-11 14:14:48 -07:00
A. Unique TensorFlower
65eadd3a02 Enable parallel processing in the Scikit-Learn models. 
Add support for configuring the parallel processing backend for Scikit-Learn while setting up the attack models.

PiperOrigin-RevId: 446844669
 2022-05-05 16:21:15 -07:00
A. Unique TensorFlower
7eea74a6a1 COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/privacy/pull/230 from npapernot:hyperparam 8835b9c4072e3e598aa49d605e7643a2c2e65988 
PiperOrigin-RevId: 446832781
 2022-05-05 15:38:47 -07:00
A. Unique TensorFlower
930c4d13e8 Handle the case when the data comes from a multilabel classification problem but the provided samples happen to have just one positive label per sample. 
PiperOrigin-RevId: 445468067
 2022-04-29 11:36:47 -07:00
Galen Andrew
81d5880702 Migrate accounting in tutorials to Google DP. 
PiperOrigin-RevId: 444993855
 2022-04-27 16:09:27 -07:00
Fabien Hertschuh
d47cc695cd Consistently import all estimator symbols via tensorflow_estimator as opposed to using tensorflow.estimator for some symbols. 
PiperOrigin-RevId: 443417581
 2022-04-21 10:36:59 -07:00
Galen Andrew
868cf54470 Deprecates implementations of RDP accounting from tensorflow_privacy in favor of differential_privacy. 
PiperOrigin-RevId: 443177278
 2022-04-20 13:25:48 -07:00
A. Unique TensorFlower
ee35642b90 Add multi-label support for Tensorflow Privacy membership attacks. 
PiperOrigin-RevId: 443176652
 2022-04-20 13:23:26 -07:00
Galen Andrew
e14618fe7c Remove DpEvent from tensorflow_privacy which has been migrated to Google DP libraries. 
PiperOrigin-RevId: 441490619
 2022-04-13 08:40:52 -07:00
Galen Andrew
34ad9dbedc Automated rollback of commit 7e89dad685 
PiperOrigin-RevId: 441245348
 2022-04-12 11:34:14 -07:00
A. Unique TensorFlower
7e89dad685 Delete TF Privacy fork of the Google DP accounting API. 
PiperOrigin-RevId: 440992901
 2022-04-11 14:12:46 -07:00
A. Unique TensorFlower
34f8774dad Migrate more references to dp_event to Google DP libraries. 
PiperOrigin-RevId: 440230682
 2022-04-07 16:39:38 -07:00
A. Unique TensorFlower
676945d3a6 Migrate references to dp_event to Google DP libraries. 
PiperOrigin-RevId: 440204574
 2022-04-07 14:42:57 -07:00
Fabien Hertschuh
5493a3baf0 Explicitly import estimator from tensorflow as a separate import instead of 
accessing it via tf.estimator and depend on the tensorflow estimator target.

PiperOrigin-RevId: 438419860
 2022-03-30 16:05:01 -07:00
Fabien Hertschuh
fc2c15ab21 Explicitly import estimator from tensorflow as a separate import instead of 
accessing it via tf.estimator and depend on the tensorflow estimator target.

PiperOrigin-RevId: 437818180
 2022-03-28 12:00:58 -07:00
Steve Chien
70ab071e23 In dp_optimizer_keras.py, set the default value of unconnected gradients to be 'zero'. 
PiperOrigin-RevId: 435759367
 2022-03-18 16:10:24 -07:00
A. Unique TensorFlower
d21e492be6 [Py Accounting] Add typing annotations in RDP accounting. 
PiperOrigin-RevId: 435703861
 2022-03-18 12:20:03 -07:00
A. Unique TensorFlower
adde2064dd Remove dependence on six and use super() in rdp_privacy_accountant 
PiperOrigin-RevId: 434668652
 2022-03-14 23:24:33 -07:00
A. Unique TensorFlower
42df23eb79 Add Laplace DpEvent 
PiperOrigin-RevId: 432475405
 2022-03-04 10:14:36 -08:00
Shuang Song
2c65cc7910 In binary log loss for membership inference attack, allow prediction to have shape (n, 1). 
PiperOrigin-RevId: 432267275
 2022-03-03 13:17:49 -08:00
Shuang Song
767788e9cf Adds per-example membership scores to trained attackers. 
PiperOrigin-RevId: 431615160
 2022-02-28 23:52:21 -08:00
Steve Chien
a33afde0c1 Add ability to specify number of microbatches in DPModel class. 
PiperOrigin-RevId: 430358084
 2022-02-22 20:32:05 -08:00
Steve Chien
bfdcb7f64f Update version number to 0.8.0. 
PiperOrigin-RevId: 430268000
 2022-02-22 12:19:05 -08:00
Shuang Song
ec7d44237c Allow customized loss functions for membership inference attack. 
PiperOrigin-RevId: 430267951
 2022-02-22 12:17:56 -08:00
Peter Kairouz
39fa1d361f Add DistributedSkellamSumQuery symbol to public API. 
PiperOrigin-RevId: 430261878
 2022-02-22 11:52:54 -08:00
Shuang Song
12541c23d4 Fix MIA readme: labels are not needed in basic usage. 
PiperOrigin-RevId: 430230630
 2022-02-22 09:37:42 -08:00
Shuang Song
7d5a57f0a8 An example for running secret sharer on image classification model. 
PiperOrigin-RevId: 430083697
 2022-02-21 14:04:51 -08:00
Shuang Song
04dd758c8a In secret generation for secret sharer, use np.random.RandomState. Restructure generate_secrets. 
PiperOrigin-RevId: 430082580
 2022-02-21 13:54:09 -08:00
Peter Kairouz
89de03e0db Adds DistributedSkellamQuery to public TF Privacy. 
PiperOrigin-RevId: 429664212
 2022-02-18 15:48:20 -08:00
Michael Reneer
69d84d1892 Add TensorFlow Privacy BUILD and WORKSPACE files. 
PiperOrigin-RevId: 429141704
 2022-02-16 23:30:06 +00:00
Shuang Song
11b4c318a2 Add links to secret sharer colab. 
PiperOrigin-RevId: 428844473
 2022-02-16 23:27:32 +00:00
Francois Chollet
085b7ddfec Remove deprecated TF1 Layer APIs apply(), get_updates_for(), get_losses_for(), and remove the inputs argument in the add_loss() method. 
PiperOrigin-RevId: 428134172
 2022-02-11 18:35:49 -08:00
Shuang Song
560926ea22 Functions for advanced MIAs. 
PiperOrigin-RevId: 428111799
 2022-02-11 16:06:47 -08:00
Francois Chollet
cd38172f35 Remove usage of deprecated layer method. 
PiperOrigin-RevId: 427339701
 2022-02-08 17:45:45 -08:00
Michael Reneer
bb153c1f9d Internal 
PiperOrigin-RevId: 427327078
 2022-02-08 16:44:37 -08:00
Michael Reneer
28db674240 Ensure that TF 1.0 API is referenced at the call site in TensorFlow Privacy. 
This change makes it easy to search for usage of TF 1.0 API and updates the TF imports across TFP to be written consistently.

PiperOrigin-RevId: 427043028
 2022-02-07 16:06:22 -08:00
Shuang Song
ceced43d0b For seq2seq MIA test, call threshold attacker directly. 
PiperOrigin-RevId: 426941426
 2022-02-07 09:46:03 -08:00
Zheng Xu
2fe51d2eeb Minor fix of tree aggregation accounting docstring. 
PiperOrigin-RevId: 426398939
 2022-02-04 07:32:22 -08:00
Michael Reneer
c3e7f08fa5 Update to import TF 2.0 in many TensorFlow Privacy Python modules. 
PiperOrigin-RevId: 425901093
 2022-02-02 09:17:30 -08:00
Michael Reneer
778c804d1b Delete unused TF 1.0 API in TensorFlow Privacy. 
PiperOrigin-RevId: 425900761
 2022-02-02 09:16:06 -08:00
Zheng Xu
6fde7b0480 Add clearer connection of TFP tree aggregation query and accounting arguments. 
PiperOrigin-RevId: 425790945
 2022-02-01 21:34:37 -08:00
Michael Reneer
8a6827b27c Update to use TF 2.0 API in TensorFlow Privacy: 
tf.logging -> Removed for absl
tf.assert_type -> tf.debugging.assert_type
tf.assert_less_equal -> tf.debugging.assert_less_equal
tf.global_norm -> tf.linalg.global_norm

PiperOrigin-RevId: 425730344
 2022-02-01 15:29:09 -08:00
Shuang Song
438da5a09b For secret sharer exposures, allow more dictionary key types, and break ties for the same perplexities. Fix a bug in the test. 
PiperOrigin-RevId: 425446829
 2022-01-31 13:27:55 -08:00
Michael Reneer
fd242e76b9 Internal 
PiperOrigin-RevId: 425430059
 2022-01-31 12:17:02 -08:00
Michael Reneer
36b8ea34ef Internal 
PiperOrigin-RevId: 424965569
 2022-01-28 15:01:34 -08:00
Michael Reneer
25571aaf87 Fix up remaining lint and format in TensorFlow Privacy. 
PiperOrigin-RevId: 424948247
 2022-01-28 13:44:12 -08:00
Michael Reneer
b37aef1751 Fixed some lint errors in TensorFlow Privacy. 
* Fixed `g-backslash-continuation`
* Fixed `g-generic-assert`
* Fixed `g-generic-assert`
* Fixed `raise-missing-from`
* Fixed `unused-argument`

PiperOrigin-RevId: 424931881
 2022-01-28 12:31:00 -08:00
Michael Reneer
e6536597c5 Fixed some lint errors in TensorFlow Privacy. 
* Fixed `g-importing-member`
* Fixed `g-bad-import-order`

PiperOrigin-RevId: 424926847
 2022-01-28 12:10:05 -08:00
Michael Reneer
943ef91ee9 Format TensorFlow Privacy files. 
PiperOrigin-RevId: 424923635
 2022-01-28 11:57:24 -08:00
Michael Reneer
07230a161a Internal 
PiperOrigin-RevId: 424922009
 2022-01-28 11:50:35 -08:00
Michael Reneer
7396ad62da Update TensorFlow Privacy to use Python 3 super(). 
PiperOrigin-RevId: 424916118
 2022-01-28 11:26:34 -08:00
Michael Reneer
9050f18b59 Update TensorFlow Privacy to use Python 3 metaclass. 
PiperOrigin-RevId: 424773127
 2022-01-27 20:32:04 -08:00
Michael Reneer
b0803999ad Add license and missing build targets to __init__.py files in TensorFlow Privacy. 
* Added license.
* Removed `.../membership_inference_attack/codelabs/__init__.py`, this code does not look like it needs to be part of the Python package.
* Removed `.../research/pate_2017/__init__.py`, this code does not look like it needs to be part of the Python package.

PiperOrigin-RevId: 424682641
 2022-01-27 12:40:23 -08:00
Michael Reneer
c36ce6d799 Normalize mpmath imports in TensorFlow Privacy to be more friendly with strict dependencies and lint. 
PiperOrigin-RevId: 424681602
 2022-01-27 12:35:43 -08:00
Michael Reneer
81a11eb824 Remove the mock dependency from TensorFlow Privacy, this is now part of the Python standard library. 
PiperOrigin-RevId: 424681527
 2022-01-27 12:34:59 -08:00
Michael Reneer
a749ce4e30 Remove TensorFlow Privacy __future__ imports. 
Note: Exclude changes to the research directory.
PiperOrigin-RevId: 424650953
 2022-01-27 10:37:22 -08:00
Michael Reneer
cfb1b881d8 Normalize scipy imports in TensorFlow Privacy to be more friendly with strict dependencies and lint. 
PiperOrigin-RevId: 424649853
 2022-01-27 10:33:23 -08:00