Added implementation of https://eprint.iacr.org/2024/1077

[sandy9999]

Added implementation of the Decision Tree protocol from "Securely Training Decision Trees Efficiently" by Divyanshu Bhardwaj, Sandhya Saravanan, Nishanth Chandran and Divya Gupta : eprint link.

This decision tree training protocol has communication complexity $O(mN \log N + hmN + hN \log N)$ which is an improvement of $\approx min(h, m, \log N)$ over Hamada et. al link.

-Sandhya Saravanan

[mkskeller]

Thank you for submitting this pull request. However, the code isn't suitable for inclusion due to the following reasons:

• Using the code fails with an indentation error
• If the error is corrected, it fails because radix_sort_permutation_from_matrix is not available
• I don't see a good reason for a separate module if the functionality it provides is essentially the same as an existing module
• decision_tree_new is not a meaningful name and will be even less meaningful as time goes on
• Multiple functions are just copy-pasted instead of imported. This is bad practice as future improvements to them will go unused.

I would suggest the following apart from correcting the compilation errors:

• Add the code to the decision_tree module where it can be found by users.
• Avoid using "new" as prefix or suffix. Use either "optimized" or something more prescriptive.
• Avoid the use of entirely copy-pasted functions.
• Consider sub-classing the existing classes.

[mkskeller]

Thank you for your additions. I'm afraid I still see two issues here:

1. Running Scripts/compile-emulate.py breast_tree nearest degrades between the main branch and this. This might be related to the fact that the main branch outputs a leave NID of 32 but in this branch the NID's only go up to 16.
2. I'm hesitant to merge this branch into the main branch because of the separate module. The reason is that, in the long term, I want to incorporate the changes into the main functionality in the decision_tree module. After that, anyone who used the decision_tree_optimized module will need to change their code as I don't to add any hacks. If your functionality would be available within decision_tree module the adaption would be much more smooth. That said, I'm happy to incorporate your changes in a separate branch, and I indent to merge that in the future to give you credit. Please let me know your thoughts on this.

[mkskeller]

After having another look, the first issue might be related to not calling UpdateState after the last layer.

[sandy9999]

Hi @mkskeller Thank you for your detailed comments. I understand your concerns, and will try to work with the decision_tree module.

[sandy9999]

Hi @mkskeller I have updated the PR as discussed. Would this work?

[mkskeller]

I'm afraid there are still a few issues:

1. decision_tree_optimized still exists, did you mean to delete this?
2. The changes to decision_tree break all applications thereof, adult, easy_adult, bench-dt, and breast_tree (if changed to decision_tree).

[sandy9999]

Hi @mkskeller

Sorry for the extremely late follow up, got missed. 1 and 2 have been addressed. That is,

./compile.py -Z 3 -R 64 breast_tree
(tmux)
./replicated-ring-party.x 0 breast_tree
./replicated-ring-party.x 1 breast_tree
./replicated-ring-party.x 2 breast_tree

and

./compile.py -Z 3 -R 64 bench-dt 10 10
(tmux)
./replicated-ring-party.x 0 bench-dt-10-10
./replicated-ring-party.x 1 bench-dt-10-10
./replicated-ring-party.x 2 bench-dt-10-10

do not break anymore. adult and easy_adult broke even on the older version of decision_tree.py (with adult, I need instructions on how to take input, and for easy_adult the process got killed).