Quick feedback

[dimakuv]

2. The Coordinator performs attestation on the TensorFlow Serving application running on SGX using Graphene, and provides the application with secrets defined by the Marblerun manifest.
3. The Administrator encrypts a pre-trained model and uplodas it to Kubernetes file storage

I would swap these two steps. It makes more logical sense to first upload the model to the storage, and then spawn & attest the TensorFlow app. (Also should change the numbering on the diagram.)

3. The Administrator encrypts a pre-trained model and uplodas it to Kubernetes file storage

Typo: uploads.

Get Marbleruns intermediate certificate to connect to the model server

Typo: Marbelrun's.

https://github.com/edgelesssys/graphene-tensorflow-demo/blob/master/tensorflow.dockerfile#L62

You seem to use a slightly outdated version of Graphene. We now moved (well, in the process of moving) to the Meson build & install. In particular, Graphene is now able to install itself (only the needed binaries and related files) either under /usr/local/ (default path) or in your specific path via --prefix. See https://graphene.readthedocs.io/en/latest/building.html#id1 and especially https://graphene.readthedocs.io/en/latest/building.html#additional-build-options.

I didn't check the non-Graphene files and commands (like Kubernetes files), but overall looks correct. Thanks for this!