Bug on Zeroes Multiplication #308
Labels
bug
Something isn't working
Comments
|
Thanks for pointing this out. I am able to reproduce this. This only happens in the 3+ party setting implying it may be a result of the division protocol. However, I'm not sure why this would happen in this case. Looking into it now. |
knottb
added a commit
to knottb/CrypTen-2
that referenced
this issue
Sep 27, 2021
Summary: Fixes bug described in facebookresearch#308 This bug causes wrap-around errors in 3+ party division. Upon inspection, it became clear that this bug is caused by a change in the scaling logic introduced in D29064976 (facebookresearch@a632bf4). This bug happened because the correction applied to wraps has different scaling than the input tensor, and the new logic rescales accordingly. This can be fixed by applying the arithmetic on the shares rather than the ASTs. To ensure bugs like this won't be reproduced again, I've introduced 3PC tests into test_mpc.py as well. These will increase runtime of the test harness, making it more necessary to upgrade the testing framework for easier submissions. Differential Revision: D31206649 fbshipit-source-id: b7fdc92daebe169d8a95ff8bcb131391ceb831e8
This was referenced Sep 27, 2021
facebook-github-bot
pushed a commit
that referenced
this issue
Sep 28, 2021
Summary: Pull Request resolved: #313 Pull Request resolved: fairinternal/CrypTen#243 Fixes bug described in #308 This bug causes wrap-around errors in 3+ party division. Upon inspection, it became clear that this bug is caused by a change in the scaling logic introduced in D29064976 (a632bf4). This bug happened because the correction applied to wraps has different scaling than the input tensor, and the new logic rescales accordingly. This can be fixed by applying the arithmetic on the shares rather than the ASTs. To ensure bugs like this won't be reproduced again, I've introduced 3PC tests into test_mpc.py as well. These will increase runtime of the test harness, making it more necessary to upgrade the testing framework for easier submissions. Reviewed By: shubho Differential Revision: D31206649 fbshipit-source-id: 59e85ac6f7fb07591aa55fda1181eb3e3a9aeb00
|
This should be fixed in #313. Closing this issue. |
|
Thanks a lot for fixing the issue! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
On multiplying two zero tensors from source 0 in 5 party setting, I was getting large +/- values (4.2950e+09). I have also attached the jupyter notebook containing the code and requirements.txt file containing dependencies information.
Jupyter Notebook
Python: 3.7.11
Dependencies
Please help. Thanks in advance.
The text was updated successfully, but these errors were encountered: