Open neha-jawalkar opened 2 years ago
Hey Neha,
Just wanted to drop you a quick update about this since that email thread a week or two ago -- I haven't forgotten about this issue! Just very underwater, so hopefully I should be able to push something in the next few days.
Thanks for the update!
Hello,
Insofar as I can tell, in the 2PC code, shares of beaver triples are generated as (0, 0, 0) for both parties. Due to this (and some other code), shares of all intermediate values are of the form (X, 0). This is worrying for two reasons. First, this is insecure and may produce incorrect latency numbers. Second, all truncations are in fact exact and do not emulate local truncation errors in MPC when done with correct beaver triples. Could you suggest a way to measure accuracy when the MPC suffers from errors caused by local truncation?
Thanks!