I think that this RAM is a consequence of the library / scheme used which is SEAL - CKKS, and I would like to know if the RAM is needed per client or a single one is needed ? I mean, on a single server, if I have two clients would I need 2 times the RAM needed for one ?
In the case of Cryptonets / LoLa , if I have two different clients that want to infer a CIFAR10 image, would I need 2 * 12GB to infer the two of them ? Or only 12GB is needed ?
Hi everyone,
First thank you for your awesome lib.
I had a question about the RAM needed to use SEAL per client.
I have seen in some works: Low-Complexity Deep Convolutional Neural Networks on Fully Homomorphic Encryption Using Multiplexed Parallel Convolutions or Low Latency Privacy Preserving Inference that a lot of RAM is needed (resp. 512GB and 12GB for each of these works). I have tried to open issues on their github but unfortunately it is not possible anymore.
I think that this RAM is a consequence of the library / scheme used which is SEAL - CKKS, and I would like to know if the RAM is needed per client or a single one is needed ? I mean, on a single server, if I have two clients would I need 2 times the RAM needed for one ? In the case of Cryptonets / LoLa , if I have two different clients that want to infer a CIFAR10 image, would I need 2 * 12GB to infer the two of them ? Or only 12GB is needed ?
Thanks in advance for your answer