4 ms·

I am not an encryption expert, but to the casual techy follower, homomorphic encryption seems like it could be huge for the union of machine learning and digita

I am not an encryption expert, but to the casual techy follower, homomorphic encryption seems like it could be huge for the union of machine learning and digital privacy. Are there any experts on HN who could comment on how mature this technology is? Or does anyone know of any interesting projects that use this? I know Numerai is using something of this sort, but it is hard for me to tell how legitimate it is.

rtkwe 8y ago
Another big advantage I'd hope to see is it enabling more interesting volunteer computing efforts or maybe a truly distributed web infrastructure.
osaariki 8y ago
I work at Microsoft Research with the SEAL team on applications for Fully Homomorphic Encryption (FHE). Batched integer FHE schemes are great for machine learning. The regular data access patterns in neural networks make using the batching easy and the rather "wide" computations lead to good encryption parameters. Also a new scheme, CKKS, gives the ability to divide by certain plaintext constants, which helps control scaling in fixed point representations. This comes in exchange for results being approximate, which is not a problem in machine learning. Currently what FHE needs is better toolchain support. Selecting encodings and encryption parameters is hard, but a lot can be done with good language support and tooling. We're presenting some of our work on automatically selecting good data layouts for NNs on FHE at the PPML workshop in NIPS next weekend [1]. [1] https://ppml-workshop.github.io/ppml/ https://ppml-workshop.github.io/ppml/
dboreham 8y ago
One issue to be aware of : it is typically very (very) slow.
garganzol 8y ago
Homomorphic encryption (full and partial) opens the road to some unforeseen forms of obfuscation. This basically means that we'll see the status quo change in this area quite soon. While the full homomorphic encryption schemes tend to be at the slower side for now, partially homomorphic systems are fast.

rtkwe 8y ago

Another big advantage I'd hope to see is it enabling more interesting volunteer computing efforts or maybe a truly distributed web infrastructure.

osaariki 8y ago

I work at Microsoft Research with the SEAL team on applications for Fully Homomorphic Encryption (FHE). Batched integer FHE schemes are great for machine learning. The regular data access patterns in neural networks make using the batching easy and the rather "wide" computations lead to good encryption parameters. Also a new scheme, CKKS, gives the ability to divide by certain plaintext constants, which helps control scaling in fixed point representations. This comes in exchange for results being approximate, which is not a problem in machine learning. Currently what FHE needs is better toolchain support. Selecting encodings and encryption parameters is hard, but a lot can be done with good language support and tooling. We're presenting some of our work on automatically selecting good data layouts for NNs on FHE at the PPML workshop in NIPS next weekend [1]. [1] https://ppml-workshop.github.io/ppml/ https://ppml-workshop.github.io/ppml/

dboreham 8y ago

One issue to be aware of : it is typically very (very) slow.

garganzol 8y ago

Homomorphic encryption (full and partial) opens the road to some unforeseen forms of obfuscation. This basically means that we'll see the status quo change in this area quite soon. While the full homomorphic encryption schemes tend to be at the slower side for now, partially homomorphic systems are fast.