Fully Homomorphic Encryption and risc zero

[GuangguangLu]

I want to combine risc zero and FHE, but I haven't found a suitable fully homomorphic encryption tool yet. Do you know whichcan be combined with risc zero?

[GuangguangLu]

I tried many rust FHE libraries, but none of them worked, is my approach wrong? Are there any successful projects?
I tried:

https://crates.io/crates/sunscreen

[SchmErik]

It looks like this one uses the risc zero zkvm. https://github.com/emilianobonassi/zkFHE Aside from this we are not aware of any FHE projects that use risc zero or an FHE library that runs on risc zero

[GuangguangLu]

I saw this project, but it seems that it only supports addition, not full homomorphic encryption

[emilianobonassi]

Thanks, @SchmErik for mentioning it. @guangguang888 I'm happy to see someone else trying zk FHE.

I've been using addition only HE but originally wanted to go for https://github.com/zama-ai/tfhe-rs which supports FHE.

The current challenges (not actively working on that) are:

1. tfhe latest version leverage parallel processing and risc0 is 'single core' atm. I had to use a previous version and tune for not using parallel libraries i.e. https://github.com/emilianobonassi/tfhe-rs/commits/0.1.5b

2. when I tried libm support was not integrated into risc0 and I had to fork it, currently the latest release includes it but I had issues similar to Guest fails to build crossbeam due to atomics #444

3. zama fhe keys are pretty big (>100MB) and using common serialization/deserialization lead was pretty slow on the host. I've tried different paths like unsafe mem copy and casting but still had some issues. Indeed, I had to double the memory available in risc0 i.e. emilianobonassi/risc0@b871031

Using my latest version of risc0 you are just facing 3. but would be best to rebase to the latest official version and start from there.

For the logic, just split client key generation in the host and server manipulation in the guest/methods. Useful docs here https://docs.zama.ai/tfhe-rs/boolean/tutorial

Happy to support/build together @guangguang888

Arm yourself with a lot of patience because testing requires a few mins every run :D

PS: it would be nice if risc0 or zama would provide some grant on that to prioritize/incentivize this work

[GuangguangLu]

Thank you very much for your reply.
I will try patiently, thanks again for your reply.

[GuangguangLu]

Do you mean it is feasible to combine tfhe-rs and risc zero? By using a suitable version and considering the issue of tfhe key size.

[emilianobonassi]

Do you mean it is feasible to combine tfhe-rs and risc zero? By using a suitable version and considering the issue of tfhe key size.

Frankly, I don’t know. My objective with the PoC was to test it and see in the case the margins for optimization.

[SchmErik]

Thanks, @SchmErik for mentioning it. @guangguang888 I'm happy to see someone else trying zk FHE.

I've been using addition only HE but originally wanted to go for https://github.com/zama-ai/tfhe-rs which supports FHE.

The current challenges (not actively working on that) are:

1. tfhe latest version leverage parallel processing and risc0 is 'single core' atm. I had to use a previous version and tune for not using parallel libraries i.e. https://github.com/emilianobonassi/tfhe-rs/commits/0.1.5b

2. when I tried libm support was not integrated into risc0 and I had to fork it, currently the latest release includes it but I had issues similar to [Guest fails to build crossbeam due to atomics #444](https://github.com/risc0/risc0/issues/444)

3. zama fhe keys are pretty big (>100MB) and using common serialization/deserialization lead was pretty slow on the host. I've tried different paths like unsafe mem copy and casting but still had some issues. Indeed, I had to double the memory available in risc0 i.e. [emilianobonassi/risc0@b871031](https://github.com/emilianobonassi/risc0/commit/b871031da429c409710b44e4cdca1e2df1f0150d)

That's pretty big. I don't know much about FHE but is the key a single key or are there different components to the key? In other words, do you need all of the key at once or are you ok taking parts of the key?

[emilianobonassi]

Thanks, @SchmErik for mentioning it. @guangguang888 I'm happy to see someone else trying zk FHE.
I've been using addition only HE but originally wanted to go for https://github.com/zama-ai/tfhe-rs which supports FHE.
The current challenges (not actively working on that) are:

1. tfhe latest version leverage parallel processing and risc0 is 'single core' atm. I had to use a previous version and tune for not using parallel libraries i.e. https://github.com/emilianobonassi/tfhe-rs/commits/0.1.5b

2. when I tried libm support was not integrated into risc0 and I had to fork it, currently the latest release includes it but I had issues similar to [Guest fails to build crossbeam due to atomics #444](https://github.com/risc0/risc0/issues/444)

3. zama fhe keys are pretty big (>100MB) and using common serialization/deserialization lead was pretty slow on the host. I've tried different paths like unsafe mem copy and casting but still had some issues. Indeed, I had to double the memory available in risc0 i.e. [emilianobonassi/risc0@b871031](https://github.com/emilianobonassi/risc0/commit/b871031da429c409710b44e4cdca1e2df1f0150d)

That's pretty big. I don't know much about FHE but is the key a single key or are there different components to the key? In other words, do you need all of the key at once or are you ok taking parts of the key?

Idk, i’ve been using as a blackbox.

btw, relevant update just 1hr ago which reduces considerably the size for the key.

zama-ai/tfhe-rs#322