Wrong result of multiplication of a MPCTensor and a public value

[xutongye]

Description

When multiply two large (relative to ring_size) values, the reconstructed result is wrong.

As I see it, when:

• a wraparound (or say mod operation) happens in the computation process,
• and the scale of the encoder is greater than 1,

then the result will be wrong.

How to Reproduce

Please run the following code, there are two examples showing this bug.

sy.logger.remove()
import torch
from sympc.session import Session
from sympc.config import Config
from sympc.session import SessionManager
from sympc.tensor import MPCTensor
parties = [sy.VirtualMachine(name=n).get_root_client() for n in ["alice", "bob"]]

# Example 1:
session = Session(parties=parties)
SessionManager.setup_mpc(session)
x_secret = torch.Tensor([1<<30])
x = x_secret.share(session=session)
z = x * 3
print(f"expect {x_secret*3}, but get {z.reconstruct()}")

# Example 2:
session = Session(parties=parties, ring_size=256, config=Config(encoder_base=10, encoder_precision=1))
SessionManager.setup_mpc(session)
x_secret = torch.Tensor([[0.1, -1], [-4, 4]])
x = x_secret.share(session=session)
z = x * 3
print(f"expect {x_secret*3}, but get {z.reconstruct()}")