Determine whether differential privacy/encrypted learning can be applied on audio and text for TTS
Opened this issue · 1 comments
Determine whether it is possible to encrypt the audio, and the learned audio/model using an encryption to which nobody has the key, to learn on homogenously encrypted data without sharing audio data with an external server.
Also determine whether the text output can be encrypted such that only a user with a private key is able to get the text output, whilst still allowing the model to learn from the labels, (e.g. if the user corrects the text).
In this case, make the labels the actual appointment ics files after the user has manually verified and/or corrected them.
I think yes:
Paper:
https://arxiv.org/abs/2205.11935
Repo:
https://github.com/IAIK/CryptoTL
Todo, determine how homomorphic encryption is combined with differential privacy in context of transfer learning.