I want to use secure multiparty computation (SMPC) during inference in my model which is built using TensorFlow will this library integrate well with my TensorFlow model or it is just supported with pytorch?

[Am0stafa]

I would love the community to discuss with me the potential errors I might get when integrating it with TensorFlow.
Thanks

[madhavajay]

Hi @Am0stafa we had tensorflow and tensorflow federated integrated into an older version of syft, I think version 0.7.0. Since 0.8.0 we moved to supporting jax and torch as the main DL libraries, but since we now have the ability to define custom docker images for the workload you can essentially use any python library you like.

Take a look at this example notebook:
https://github.com/OpenMined/PySyft/blob/dev/notebooks/api/0.8/10-container-images.ipynb

tensorflow_dockerfile_str = f"""
FROM openmined/grid-backend:{syft_base_worker_tag}

RUN pip install tensorflow

""".strip()

tensorflow_dockerfile_str = "openmined/custom-worker-tensorflow:1.0.0"

The main thing you will need to do is make sure the version of tensorflow is compatible with syft. We install Syft in google colab so what ever version they use is probably compatible, best way to check is to try pip install tensorflow in a virtualenv with syft and see what versions resolve etc. The common dependencies are things like numpy.

Once you have your image with a working version of tensorflow make sure that the inputs and outputs of your function are supported by serialization.

For now I would recommend using numpy arrays or jax tensors as they are supported. You can see an example of constructing jax tensor weights here:
https://github.com/OpenMined/PySyft/blob/dev/notebooks/api/0.8/04-jax-example.ipynb

[madhavajay]

@Am0stafa also feel free to join us on slack if you want to chat directly:
https://slack.openmined.org/

[madhavajay]

I will close this issue for now. We will be adding more examples for DL soon but for now the above custom container workload information and notebook on how to send in weights and utilize custom class code should be sufficient.