DP-InstaHide Trainer Implementation #1956
Conversation
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Codecov Report
@@ Coverage Diff @@
## dev_1.13.0 #1956 +/- ##
==============================================
+ Coverage 84.96% 85.47% +0.51%
==============================================
Files 285 286 +1
Lines 24912 25030 +118
Branches 4477 4495 +18
==============================================
+ Hits 21166 21394 +228
+ Misses 2612 2486 -126
- Partials 1134 1150 +16
|
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
|
The implementation and unit tests are complete and ready for review, but will also be adding a notebook demo. |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
|
A review comment, please also add the new notebook to the notebook README file. |
| x_aug = self._generate_noise(x_aug) | ||
|
|
||
| # extract label reduction and set to no reduction if needed | ||
| reduce_labels = self._classifier._reduce_labels # type: ignore # pylint: disable=W0212 |
There was a problem hiding this comment.
Why is there a type ignore?
There was a problem hiding this comment.
Not all classifiers of type CLASSIFIER_LOSS_GRADIENTS_TYPE have the _reduce_labels property which causes a mypy error. This line is needed due to the PyTorchClassifier always reducing labels for nn.CrossEntropyLoss.
There was a problem hiding this comment.
Ho @f4str Thank you very much for contributing DP-InstaHide to ART!
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
Signed-off-by: Farhan Ahmed <Farhan.Ahmed@ibm.com>
There was a problem hiding this comment.
Hi @f4str Thank you very much for contributing DP-InstaHide to ART! The review changes look good to me.
Description
Implementation of the DP-InstaHide trainer. This trainer builds upon the previously implemented data augmentation defenses. Also includes a jupyter notebook demo showing how to use the trainer and its effectiveness on MNIST (a reference was added to the notebook README).
As a consequence, the following two bugs/issues were resolved
check_and_transform_label_formatfunction when calling thecompute_lossfunction in theTensorFlowV2ClassifierandKerasClassifierclasses. This fixes the argmax issue when passing integer labels. (Consolidated from PR Validate labels incompute_losscall forTensorFlowV2ClassifierandKerasClassifier#1960)_probability_labelsin thePyTorchClassifierand have it set to True when usingnn.CrossEntropyLoss. This allows thereduce_labels()function to check if the labels are one-hot encoded or a probability distribution and handle it accordingly. This allows support for probability distribution labels. (Consolidated from PR Add Support for Probability Distribution Labels in PyTorchClassifier #1967)Fixes #1827
Fixes #1959
Fixes #1966
Type of change
Please check all relevant options.
Testing
Please describe the tests that you ran to verify your changes. Consider listing any relevant details of your test configuration.
Test Configuration:
Checklist