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

139 commits

This branch
This branch
All branches
Author SHA1 Message Date
A. Unique TensorFlower
5cd2439401 Remove calls to _dp_sum_query.set_batch_size in dp_optimizer.py, as no method with that name exists for objects of class QueryWithLedger. 
PiperOrigin-RevId: 259858031
 2019-07-24 18:08:35 -07:00
A. Unique TensorFlower
28fd879864 Merge pull request #57 from jvmancuso:qadapt-nest-ver 
PiperOrigin-RevId: 259592947
 2019-07-23 12:56:31 -07:00
jvmancuso
3072c86c79 find nest module based on TF version for quantile_adaptive_clip_sum_query.py 2019-07-20 10:43:48 -04:00
Nicolas Papernot
3de1fcd829 add pylint to README 
PiperOrigin-RevId: 259029555
 2019-07-19 13:41:07 -07:00
A. Unique TensorFlower
98723b9c3a Added rdp_accountant dependency to privacy/BUILD. 
PiperOrigin-RevId: 258657061
 2019-07-17 15:44:16 -07:00
Nicolas Papernot
bb7956ed7e fix keras typo 
PiperOrigin-RevId: 258434656
 2019-07-16 13:59:14 -07:00
Galen Andrew
973a1759aa Remove unused global_state reference from initial_sample_state. 
global_state is never used in any of our existing DPQueries, and we don't have any compelling use case.

PiperOrigin-RevId: 255480537
 2019-06-27 14:37:49 -07:00
Nicolas Papernot
6171474465 harmonize analysis parameters with current DPSGD API 
PiperOrigin-RevId: 255080643
 2019-06-25 17:06:01 -07:00
Ilya Mironov
45bcb3a0e4 Adding privacy analysis to the Logistic Regression for MNIST tutorial. 
PiperOrigin-RevId: 254815428
 2019-06-24 12:50:08 -07:00
A. Unique TensorFlower
2b97c7c735 Logistic regression for mnist with new privacy analysis. 
PiperOrigin-RevId: 252743967
 2019-06-11 18:58:56 -07:00
Galen Andrew
d5dcfec745 Remove set_denominator functions from DPQuery and make QueryWithLedger easier to use. 
set_denominator was added so that the batch size doesn't need to be specified before constructing the optimizer, but it breaks the DPQuery abstraction. Now the optimizer uses a GaussianSumQuery instead of GaussianAverageQuery, and normalization by batch size is done inside the optimizer.

Also instead of creating all DPQueries with a PrivacyLedger and then wrapping with QueryWithLedger, it is now sufficient to create the queries with no ledger and QueryWithLedger will construct the ledger and pass it to all inner queries.

PiperOrigin-RevId: 251462353
 2019-06-04 10:14:32 -07:00
Galen Andrew
7636945566 Cast to ensure record of NoPrivacyAverageQuery is float for compatibility with sample_state. 
PiperOrigin-RevId: 249909614
 2019-05-24 15:28:39 -07:00
Steve Chien
15c07250a1 Add dtype=tf.int32 to TensorBuffer capacity and current size. 
PiperOrigin-RevId: 249908717
 2019-05-24 15:22:43 -07:00
Nicolas Papernot
a06bc6c99b fix imports for v1 and make the versioning more explicit through LooseVersion 
PiperOrigin-RevId: 249732562
 2019-05-23 15:57:08 -07:00
Ilya Mironov
0efb23afcb Changing initial capacity for the ledger to smaller values. (+ restoring compatibility with Python 2) 
PiperOrigin-RevId: 249292683
 2019-05-21 11:38:27 -07:00
Ilya Mironov
a94dc626b1 Specifying minimal TF version required (currently 1.13, due to dependency on the train module). 
PiperOrigin-RevId: 248809713
 2019-05-17 16:39:48 -07:00
Galen Andrew
7992006077 Add quantile_adaptive_clip_sum_query to privacy package. 
PiperOrigin-RevId: 248617353
 2019-05-16 15:56:16 -07:00
Galen Andrew
3908429796 Make DPQuery classes (almost) completely functional: the only state from the initializer that is used gets pushed into the initial_global_state. 
PiperOrigin-RevId: 248424593
 2019-05-15 16:06:37 -07:00
Steve Chien
17fefb3895 Remove tf.function annotation from quantile_adaptive_clip_sum_query.py that was causing import error. 
PiperOrigin-RevId: 248236331
 2019-05-14 16:40:55 -07:00
Galen Andrew
aaf029edad Add quantile_adaptive_clip_sum_query which dynamically adjusts the clipping norm so a specified fraction of records per sample are clipped. 
PiperOrigin-RevId: 248201320
 2019-05-14 13:35:29 -07:00
Galen Andrew
1d1a6e087a Extensions to DPQuery and subclasses. 
1. Split DPQuery.accumulate_record function into preprocess_record and accumulate_preprocessed_record.
2. Add merge_sample_state function.
3. Add default implementations for some functions in DPQuery, and add base class SumAggregationDPQuery that implements some more. Only get_noised_result is still abstract.
4. Enforce that all states and parameters used as inputs and outputs to DPQuery functions are nested structures of tensors by replacing numbers with constants and Nones with empty tuples.

PiperOrigin-RevId: 247975791
 2019-05-13 11:28:56 -07:00
Steve Chien
82852c0e71 Add comments explaining the relationship between ML terminology and DP terminology. 
PiperOrigin-RevId: 246926753
 2019-05-06 17:12:24 -07:00
Ilya Mironov
9cece21d92 Clean-up pass to eliminate warnings: replacing deprecated endpoints with recommended versions and annotating test sizes. 
PiperOrigin-RevId: 246901723
 2019-05-06 14:50:23 -07:00
Ilya Mironov
85280ab568 Removing max_samples and max_queries from the tutorial. 
PiperOrigin-RevId: 246850557
 2019-05-06 10:30:35 -07:00
Steve Chien
098940bf1b Very minor formatting change. 
PiperOrigin-RevId: 246596295
 2019-05-03 16:41:38 -07:00
Steve Chien
28639ba0a8 Allow tensor buffers to automatically resize as needed. 
PiperOrigin-RevId: 246594454
 2019-05-03 16:30:02 -07:00
Steve Chien
beb86c6e18 Update PrivacyLedger and DPOptimizer to make certain arguments optional. 
PiperOrigin-RevId: 246235646
 2019-05-01 18:07:32 -07:00
Nicolas Papernot
c09ec4c22b minor fixes to improve tf 1 and 2 compatibility 
PiperOrigin-RevId: 246008822
 2019-04-30 13:22:57 -07:00
Nicolas Papernot
febafd830d update API calls for TF2 
PiperOrigin-RevId: 245817981
 2019-04-29 14:00:40 -07:00
A. Unique TensorFlower
ab466b156c Check batch_size % microbatches = 0 and calculate privacy budget only when dpsgd is set. 
PiperOrigin-RevId: 244949900
 2019-04-23 16:40:33 -07:00
Ilya Mironov
a3e03f773e Adding a paragraph to the walk-through on how to choose RDP orders. Plus deleting empty lines in rdp_accountant.py. Fixing issue #47. 
PiperOrigin-RevId: 244467825
 2019-04-19 21:50:49 -07:00
Steve Chien
31219a5f3f Fix DP optimizers to handle gradients that are None. 
PiperOrigin-RevId: 244429987
 2019-04-19 15:07:55 -07:00
A. Unique TensorFlower
134b7d2093 Merge pull request #45 from splatonline:ledger 
PiperOrigin-RevId: 244412992
 2019-04-19 13:33:57 -07:00
Galen Andrew
28df9cf233 Add missing deps to privacy __init__.py. 
PiperOrigin-RevId: 244249099
 2019-04-18 13:39:09 -07:00
Tim Garnsey
aeb6a94b59 added a ledger during optimizer instantiation to the language model tutorial 2019-04-17 12:52:08 +10:00
Ilya Mironov
51e29667d9 Fixing issue #44 (imports in privacy/__init__.py). Added __init__.py to the dp_query directory for Python 2 compatibility. 
PiperOrigin-RevId: 243329997
 2019-04-12 14:10:48 -07:00
Ilya Mironov
3c4409d6d6 Restoring ability to run compute_dp_sgd_privacy.py as a standalone script. 
At present, the script has no heavy dependencies except for the rdp_accountant, which is by itself pretty light-weight. However, importing rdp_accountant triggers __init__.py in third_party/py/tensorflow_privacy/privacy, which loads TF and all of tf.privacy. The CL adds a check to the __init__.py, which controls this behavior.

PiperOrigin-RevId: 243172355
 2019-04-11 17:06:53 -07:00
Galen Andrew
e8113a0365 Add DummyOptimizer to top-level imports. 
PiperOrigin-RevId: 242715034
 2019-04-09 12:14:22 -07:00
Steve Chien
3dc2b398d2 Closes #41 
PiperOrigin-RevId: 241786994
 2019-04-03 12:48:25 -07:00
Galen Andrew
9106a04e2c Use PrivacyLedger for privacy accounting. 
Prior to this change the PrivacyLedger is running to keep a log of private queries, but the ledger is not actually used to compute the (epsilon, delta) guarantees. This CL adds a function to compute the RDP directly from the ledger.

Note I did verify that the tutorial builds and runs with the changes and for the first few iterations prints the same epsilon values as before the change.

PiperOrigin-RevId: 241063532
 2019-03-29 15:31:32 -07:00
Nicolas Papernot
8507094f2b add walkthrough MD 
PiperOrigin-RevId: 241016765
 2019-03-29 11:24:36 -07:00
A. Unique TensorFlower
a8e54d928c Merge pull request #37 from ShadiRahimian:patch-1 
PiperOrigin-RevId: 240601653
 2019-03-27 11:13:33 -07:00
Nicolas Papernot
01e7cac7b5
Update compute_dp_sgd_privacy.py 2019-03-27 09:22:58 -07:00
Nicolas Papernot
8db2dd6bca
Update compute_dp_sgd_privacy.py 2019-03-27 09:18:45 -07:00
Nicolas Papernot
03050adb27 adding scratch for walkthrough blog post 
PiperOrigin-RevId: 240425310
 2019-03-26 14:16:27 -07:00
Shadi Rahimian
0a0e5cb3c3
line 83 produces error 
TypeError: can only concatenate list (not "range") to list
 2019-03-26 15:28:09 +01:00
Galen Andrew
6231d0802d Cleanup directory structure, add top-level imports and add normalized_query. 
Moved query classes from dir optimizers into new dir dp_query. Added NormalizedQuery class for queries that divide the output of another query by a constant like GaussianAverageQuery.

PiperOrigin-RevId: 240167115
 2019-03-25 10:21:04 -07:00
Nicolas Papernot
3c1e9994eb fix vector loss issue with Keras by instantiate a loss object 
PiperOrigin-RevId: 239483918
 2019-03-20 15:08:58 -07:00
Nicolas Papernot
0ebd134d99 Closes #33 
PiperOrigin-RevId: 239129202
 2019-03-18 22:42:01 -07:00
Galen Andrew
0aad84ab3f Move mpmath dependency to unittest target that uses it, and explicit import of mpmath functions to reduce size. 
PiperOrigin-RevId: 239056360
 2019-03-18 14:03:47 -07:00