Encrypting training data to securely run on a cloud platform?

Question

I'd like to securely run on cloud platforms*.
_{*like PyTorch on Amazon Web Services, Google Cloud Platform, Microsoft Azure, or Lightning Studios}

Are there cryptographic methods (e.g., in PyTorch) to encrypt training data; and then, after I download the model parameters to my local machine, do inference locally and decrypt the data the model outputs?

^{My question is similar to: "How do I encrypt and decrypt my model when I run inference on it?", but I'm asking about encrypting/decrypting training and inference data, not the model parameters.}

Answer 1

There are a handful of projects in the PyTorch ecosystem that help with data privacy.

PyTorch Pocket Reference pp. 228-9:

---

Table 8-8. Security and privacy projects

Project	Description
AdverTorch	Modules for adversarial examples and defending against attacks
PySyft	Library for model encryption and privacy
Opacus	Library for training models with differential privacy
CrypTen	Framework for privacy preserving ML

---

---

There's also homomorphic encryption with TenSEAL.