Just for the purpose of clarifying in terms of related work: there are other approaches that do not use either homomorphic encryption techniques, nor garbled circuits. You can use secret-sharing too (and enable more general settings like more parties and certain collusion thresholds). These are less communication-heavy than GCs and achieve better computation than HE, I believe, but typically have a large amount of communication rounds, so depending on the setting you probably want to use HE-based tools.
Either way, it’s great to see this direction being improved!
Hi Daniel! Thanks for the comment!
Do you mean using secret-sharing for the classification step?
The scenario we (and the related works that we cited) considered involves a cloud server that has already trained the model in clear, using its own datasets. Then, that model is offered as a service. Clients that want to use the service have to encrypt their data, send it to the cloud to be classified, then download the assigned class.
It seems that there are not much opportunities to involve more parties here, since each time the service is used is just about one client and the server (but if you see plausible scenarios, please, let me know).
Probably a multi-party setting is more suitable for the training phase, with several agents contributing with their own data and getting a model at the end…