1
$\begingroup$

In Privacy-preserving machine learning, GC is usually used for privacy operation such as ReLU(x) where sign(x) needs to know. However, binary secret sharing also supports such computation via comparators($[x]_{encryted}$ > 0)(this paper). While compare the performance, binary secret sharing is usually way faster than garbled circuit. But why is garbled circuit still used in many related works, for example?

$\endgroup$

1 Answer 1

2
$\begingroup$

Secret-sharing MPC has low total communication, low computation, but high round complexity. It needs a round of interaction for every layer in the circuit. If you are trying to evaluate a neural network with 100 layers, then you will need 100 rounds of interaction.

Garbled circuits have higher total communication and computation, but in just 1 or 2 rounds of interaction. The number of rounds does not depend on the structure of the circuit.

So in some settings, where the circuit is deep and/or the network has high latency, garbled circuits may result in faster MPC.

$\endgroup$
4
  • $\begingroup$ Hi Mikero, thanks a lot for your explanation. What if there are arithmetic secret sharing between every two garbled circuits, do you still need 1 or 2 rounds for the entire network? Because from what I saw in Delphi(github.com/mc2-project/delphi), they create a circuit for each layer of ReLU. $\endgroup$
    – Mumon
    Oct 29, 2022 at 3:04
  • 1
    $\begingroup$ I don't know about that specific protocol, but I know that some protocols convert between GC and secret sharing to take advantage of things that GC does well and things that secret sharing does well. Every conversion requires another round of interaction. $\endgroup$
    – Mikero
    Oct 29, 2022 at 3:33
  • $\begingroup$ @Mikero I have a few questions on your answer. In a protocol like GMW, some of the gates like NOT, XOR, XNOR can be evaluated non interactively. Only AND and OR gates need interactions and OT. I understand that the theoretical communication complexity doesn't change but I think it should be mentioned. For GC now, iirc FreeXOR also exists so there should be improvements there too. The part that I cannot understand it how a Garbled (not Garbled Thruth Table) can be evaluated in constant rounds when almost every gate of the circuits needs OT. $\endgroup$
    – tur1ng
    Oct 29, 2022 at 10:16
  • $\begingroup$ In both GC and SS-based MPC, linear gates are free (free in communication and rounds). In GC, only the input wires need OT, to allow the GC evaluator to learn the garbled input. Once the evaluator has the garbled input, they can evaluate the entire circuit non-interactively. $\endgroup$
    – Mikero
    Oct 29, 2022 at 15:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.