Custom MPC function runs within online Setup but not within an offline setup followed by an online phase (MASCOT-Protocol)

[jueren]

Dear MP-SPDZ maintainers,

first of all - thank you so much for your brilliant contribution to the field of MPC-Frameworks!

So far i was able to implement, compile and run a custom mpc-function in a 3+ player (local) environment with the following protocols (incl. test-data, ssl-setup etc):

• atlas-party.x
• shamir-party.x
• mascot-party.x (all having different characteristicts with respect to computation time, rounds, communication etc.)

I was also able to run the tutorial using mascot-party.x but also using separate preprocessing and online phases:

• mascot-offline.x -p 0 -N 2 tutorial & ... and then running
• Player-Online.x -p 0 -N 2 tutorial giving the same results...

However i was not able to run the custom function (using Player-Online.x) after doing the mascot-offline.x - Setup for the function successfully (according to my judgement). I was able to start running the function using Player-Online.x and it seems to start executing (at least the print_ln() does show some output) but it seems that the execution immediately stops when the MPC-function requests secret inputs from the players (get_input_from()), triggering the following console-outputs (for each local player):

• Purging preprocessed data because something is wrong
• Player-Online.x: Processor/PrepBase.cpp:15: static std::string PrepBase::get_suffix(int): Assertion 'thread_num >= 0' failed.

Setup:

• Dockerized V0.3.6 of the Library on Mac OX Ventura (Intel)
• SSL for 3 Players: ./Scripts/setup-ssl.sh 3
• Data for 3 Players:
  • echo 1 0 0 0 > Player-Data/Input-P0-0
  • echo 0 1 0 0 > Player-Data/Input-P1-0
  • echo 0 0 0 0 > Player-Data/Input-P2-0
• Compilation of MPC-Function using Default Parameters / Default Computational Domain
• Triggering Preprocessing: mascot-offline.x -p 0 -N 3 -f MPC_FUNCTION & mascot-offline.x -p 1 -N 3 -f MPC_FUNCTION & mascot-offline.x -p 2 -N 3 -f MPC_FUNCTION (i tried with -F as well) Result: Using security parameter 40 Preprocessing might be insufficient due to unknown requirements Writing to file in Writing to file in Player-Data//3-p-128/
• Starting Online-Phase: Player-Online.x -p 0 -N 3 -f MPC_FUNCTION & Player-Online.x -p 1 -N 3 -f MPC_FUNCTION & Player-Online.x -p 2 -N 3 -f MPC_FUNCTION Result: Purging preprocessed data because something is wrong Player-Online.x: Processor/PrepBase.cpp:15: static std::string PrepBase::get_suffix(int): Assertion 'thread_num >= 0' failed.
• MPC-Function: I was unable to format code snippet accordingly, can i send by email?

What did i miss here? Thank you for your help!

Another question - but regarding comparison of (s)ints:

• sint(1).reveal() == 1
• sint(1) == sint(1) Do both variants work? Which should one prefer?

Cheers!

[mkskeller]

Thank you for raising this.

• You can attach files by renaming them to text or Python file.
• The crucial output is "Preprocessing might be insufficient due to unknown requirements". This means that the total amount of preprocessing is unknown due to a loop whose length only comes out at run-time, and therefore, the preprocessing will be insufficient.
• If you're interested in the online/preprocessing time, you can also run the full computation with -v, which will output the cost of the two phases separately.
• The comparisons are semantically different. The first one isn't secure due to the call of reveal().

[jueren]

Thank you so much for your fast, detailed and helping answer!

Regarding Preprocessing vs Online Phase: Does this mean, that for the offline-phase, the MPC-Function need to have predefined/fixed parameters, rendering the handy dynamic-compilation-feature (see: https://github.com/data61/MP-SPDZ/blob/master/Programs/Source/test_args.mpc) "useless"? Is there any possibility / workaround to use this dynamic compilation feature (or something similar) still for preprocessing?

As far as i have seen, the compilation process nor the actual computation does offer a possiblity to handle dynamic input sizes? For compilation there are compile-time-arguments that can be used, but there seems to be no possibility to add some command-line parameters (that can be used inside the MPC-Function) for running the actual computation (even though the documentation indicates [<runtime-arg>...]-parameters here: https://mp-spdz.readthedocs.io/en/latest/readme.html?highlight=runtime-arg#running-computation). I assume, there must be some theoretical limitations for this to be the reason, but i could not figure it out yet (as opposed to the number of players, that need to be known in advance in order to build the computational domains). Can you please elaborate a little more on this or guide us towards some literature?

Thank you so much!

Cheers!

[mkskeller]

The framework is geared towards fixed lengths because this allows for the best optimization. Dynamic lengths are sometimes possible, but it might lead to suboptimal execution. However, it's inherently not possible to predict the amount of preprocessing required if there is a relevant dynamic length, so the required preprocessing won't work.