Skip to content

Latest commit

 

History

History
189 lines (122 loc) · 8.1 KB

CHANGELOG.md

File metadata and controls

189 lines (122 loc) · 8.1 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.5.9]

Added

  • We now support TensorFlow 1.14.0
  • OpenMined as organizational contributor in the README
  • Links to TF Seal and Keras blog posts in the README

Changed

  • Add more detailed explanations to the Keras Classification notebooks

Fixed

  • Public division by adding reciprotal (tf.math.reciprotal) support
  • A bug in interactive truncation which could have lead to overflow
  • SecureNN's ReLU out of memory error for large tensors. Above a certain threshold (tensor size), the tensor gets splitted. ReLU is performed on each split and the results are then concatenated.
  • An edge case in tfe.convert where it couldnt't convert a model correctly using specific special ops in the graph

[0.5.8]

Added

  • Converter support for DepthwiseConv2d layer
  • tfe.keras.GlobalAveragePooling2D: now support global average pooling operation for spatial data
  • tfe.keras.GlobalMaxPooling2D: now supports global max pooling operation for spatial data
  • Added training support basic models:
    • Backpropagation for tfe.keras.layers.dense and sigmoid activation
    • tfe.keras.losses.BinaryCrossentropy: now supports binary cross-entropy loss.
    • tfe.keras.optimizers.SGD: now supports stochastic gradient descent optimizer.
    • Added compile method to tfe.keras.models.Sequential to configure the model for training.
    • Added fit method to tfe.keras.models.Sequential to trains models for a given number of epochs.
  • Bloom's example for fast linear regression
  • Add example of Keras model training with differential privacy, combined with predictions on encrypted data.

Changed

  • More detailed error message when raising error for unknown layer arguments.

Fixed

  • Explicitly use int64 for numpy in Pond protocol because int is interpreted differently on Windows (32bits) and macOS (64 bits).
  • Converter raises exception when passing empty model.

[0.5.7]

Added

  • tfe.local_computation decorator, which is now the preferred way for providing inputs/outputs to a secure computation in TFE. All of the examples have been updated with example usage.
  • tfe.keras.layers.DepthwiseConv2D -- converter support for the layer will land in the next release.
  • More ops supported by the converter, including SplitV and tf.keras.layers.BatchNormalization
  • set_weights for tfe.keras models and layers now accepts private variables as well as numpy arrays.
  • secure_model now supports batch predictions.

Changed

  • Examples now use the tfe.keras API for building models instead of lower level tfe Ops.
  • Documentation is now generated from Google-style docstrings. As a result, we are only building docs for the tfe.keras API. Docstrings for other modules will be converted to Google-style and published progressively.

Fixed

  • tfe.keras layers will now check to see if defaults have been changed from their originals in tf.keras, and surface errors whenever modified kwargs aren't supported. Some layers were failing to instantiate because these checks were too specific.

[0.5.6]

Added

  • More tfe.keras functionality to convert tf.keras model into tfe.keras:
    • set_weights
    • from_config
    • model_from_config
    • clone_model
  • Example with conversion of tf.keras model into tfe.keras.
  • Improved handling for cases where the secure random operation is not available
  • Added methods to the converter to inspect TF and Keras graphs
  • tfe.convert now supports more than 2 inputs to tfe.concat

Fixed

  • A bug in tfe.keras.layers.Batchnorm where the offset and scale conditions where inverted
  • A bug in tfe.convert where ops with multiple outputs where not handled properly
  • A bug in tfe.convert where it couldnt't convert a model correctly when there was more than one special ops in the graph

[0.5.5]

Added

  • More tfe.keras functionality, including BatchNormalization.
  • channels_last support for tfe.layers.Batchnorm
  • tfe.convert now supports conversions for Ops with multiple output tensors.

Changed

  • All test files have been moved into the tf_encrypted namespace next to their corresponding functionality in line with the TensorFlow style guide.

Fixed

  • A bug in tfe.convert where tf.split was only returning the first element of the output tensor array.
  • An ImportError for users of TF 1.14+.
  • A bug where __radd__ and __rsub__ were actually computing __add__ and __sub__.
  • An overzealous AssertionError in the tfe.serving.QueueServer.

[0.5.4]

Added

  • More tfe.keras functionality, including Sequential model

[0.5.3]

Added

  • First steps for the new tfe.keras module closely matching that of TensorFlow.
  • More examples around private prediction and private training.
  • Various notebooks, from getting started to in-depth debugging.
  • Decoupled triple generation for Pond and SecureNN, allowing triples to be generated up front or simultaneously with other computations.

Changed

  • All code is now following the style guide of TensorFlow.

[0.5.2]

Migration to third party organization, including on GitHub and Docker Hub.

[0.5.1]

Added

  • We now support TensorFlow 1.13.1. We plan to support to 2.0+ in a future release.
  • Am example of performing secure aggregation for federated learning is now available in the examples/federated-learning/ directory.
  • New scripts in bin/ for launching TF Encrypted jobs on local machines and cloud clusters. We have primarily used these in concert with GCP & GKE.
  • The tensor seeding from 0.4.0 has been implemented for all BackingTensor types and protocols.
  • A variety of TF Ops have been added to the Converter, including an example of a "special op" (required_space_to_batch_paddings) that converts an entire subgraph of an imported Tensorflow graph into TFE instead of converting nodes in the original TensorFlow graph one-by-one. The converter will be generalized and documented much more in a future release.
  • New ops include division between a private numerator and public denominator, as well as a secure version of tf.negative.
  • The PrivateModel abstraction can now be tagged.

Breaking

  • All examples have been updated. Most now use the higher-level Keras interface to TensorFlow before building models in TF Encrypted.
  • The various BackingTensor types have been refactored and unified under a set of factories. This affects how Configs and Protocols are instantiated. Please see the examples for more.

Changed

  • The tf_encrypted.layers.Conv2D layer now automatically includes a bias add. The bias is initialized as a zero tensor if none is provided. This will be handled more elegantly in a future release.

[0.4.0]

Added

  • SecureNN with int64 support has landed, a tensorflow build with int64 matmul support must be used. See details on how to do that here
  • Cryptographically secure random numbers feature has been implemented.
  • We now send random seeds from the crypto-producer instead of full tensors when masking and computing multiplications. This reduces the amount of data sent across the wire between parties effecting operations such as mul, matmul, conv2d, and square.
  • Four new operations are now supported: Pad, BatchToSpaceND, SpaceToBatchND, ArgMax

Changed

  • There are now separate wheels published to PyPI for MacOS and Linux.
  • Various documentation updates.

[0.3.0]

Breaking

  • Default naming of crypto producer and weights provider have been changed to crypto-producer and weights-provider respectively.

Changed

  • Various documentation updates.

[0.2.0]

Breaking

  • Import path renamed from tensorflow_encrypted to tf_encrypted

Added

  • Added the beginnings of documentation which is available on readthedocs
  • Added a CHANGELOG.md file to the project root.