I was thinking we could do something similar to what federated learning does: part of the calculation is done on the client, part in the provider. With machine learning, the neural net is created by the client first and then sent to the server such that the server cannot really understand the raw data anymore.
Discussion
That’s an idea worth exploring. Although I don’t think I understand how we would turn it into a ML problem, or how the raw data would be encoded in a way that the server can’t understand it. Probably I need to learn more about ML.
Or maybe in our case, the client just needs to finish the calculation that started on the server. 🤔 Heavier, but it could work
the only way i could envision this data being ok to publish is if it involved some kind of magic cryptographic mathematics to encode the data in such a way you can verify it without revealing very much of the graph itself, but enough that you can trust the modified weightings it creates. but honestly, it's beyond me, the hipster vibes around ZKP and privacy coins have kept me from being curious to learn, up to this point.
I have been reading about federated learning these days, there's very interesting possibilities.