Deep Learning is basically doing lots of addition and multiplication; We have algorithms that allow these operations on encrypted data, without the need to decrypt the data nor to have the key to decrypt the data. So by combining the two things we can do deep learning on homomorphically encrypted data and learn meaningful things without ever looking at what the data actually is.
I agree. IP protection and data privacy issues are a better short-term use case... and fortunately we have some time to make better HE algos before any of our AIs are really getting that smart. :)
Yes, but what this is is encrypting the network instead of the data. This way, when we improve the network iteratively to reduce predictive error, we can perform all the relevant calculations homomorphically.
