M&M: Secure Two-Party Machine Learning through Efficient Modulus Conversion and Mixed-Mode Protocols

Secure two-party machine learning has made substantial progress through the use of mixed-mode protocols, but existing approaches often suffer from efficiency bottlenecks due to inherent mismatch between optimal domains of various cryptographic primitives. In response to these challenges, we introduce framework M&M, which features an efficient modulus conversion protocol. This breakthrough enables seamless integration of the most suitable cryptographic subprotocols within their optimal modulus domains with a minimal modulus conversion overhead. We further establish new benchmarks and practical optimizations for the performance of fundamental primitives, namely comparison and multiplication, across various two-party techniques.By incorporating these techniques, M&M demonstrates significant performance enhancements over state-of-the-art solutions.

This repo contains a proof-of-concept implementation for our paper M&M.

Repo Directory Description

• include/ Contains implementation of M&M's linear protocols.
• SCI/ A fork of CryptFlow2's SCI library and contains implementation of M&M's proposed protocols.
• networks/ Auto-generated cpp programs that evaluate some neural networks.
• pretrained/ Pretrained neural networks and inputs.
• patch/ Patches applied to the dependent libraries.
• credits/ Licenses of the dependencies.
• scripts/ Helper scripts used to build the programs in this repo.

Requirements

• openssl
• c++ compiler (>= 8.0 for the better performance on AVX512)
• cmake >= 3.13
• git
• make
• OpenMP (optional, only needed by CryptFlow2 for multi-threading)

Building Dependencies

• Run bash scripts/build-deps.sh which will build the following dependencies

  • emp-tool We follow the implementation in SCI that using emp-tool for network io and pseudo random generator.
  • emp-ot We use Ferret in emp-ot as our VOLE-style OT.
  • Eigen We use Eigen for tensor operations.
  • SEAL We use SEAL's implementation for the BFV homomorphic encryption scheme.
  • zstd We use zstd for compressing the ciphertext in SEAL which can be replaced by any other compression library.
  • hexl We need hexl's AVX512 acceleration for achieving the reported numbers in our paper.

• The generated objects are placed in the build/deps/ folder.

• Build has passed on the following setting

  • MacOS 11.6 with clang 13.0.0, Intel Core i5, cmake 3.22.1
  • Red Hat 7.2.0 with gcc 7.2.1, Intel(R) Xeon(R), cmake 3.12.0
  • Ubuntu 18.04 with gcc 7.5.0 Intel(R) Xeon(R), cmake 3.13
  • Ubuntu 20.04 with gcc 9.4.0 Intel(R) Xeon(R), cmake 3.16.3

Building Demo

• Run bash scripts/build.sh which will build 6 executables in the build/bin folder
  • resnet50
  • sqnet
  • densenet121

Run Demo

• Run bash scripts/sqnet-run.sh to run on the SqueezeNet model.
• Run resnet-run.sh to run on the ResNet50 model.
• Run densenet-run.sh to run on the DenseNet121 model.

The program will invoke two process, a client and a server.

You can change the SERVER_IP and SERVER_PORT defined in the scripts/common.sh to run the demo remotely. Also, you can use our throttle script to mimic a remote network condition within one Linux machine, see below.

Micro-Benchmarks

1. Run scripts/ring_to_field_test-run.sh bw to run the conversion of ring to field. Current supported bitwidth : {16, 32, 40, 48, 56, 60, 64}.

2. Run scripts/field_to_ring_test-run.sh bw to run the conversion of field to ring. Current supported bitwidth : {16, 32, 40, 48, 56, 60, 64}.

3. Run scripts/ring_extension_test-run.sh bw to run the ring extension. Base ring = $2^{64}$, target ring = $2^{64 + \text{bw}}$ Current supported lift-bitwidth : [1, 64].

4. Run scripts/bole_test-run.sh [flag0] [flag1] to run Batch Oblivious Linear Evaluation.

• Base mod and target mod are $2^l$, itermedia mod is a NTT-friendly prime (enable SIMD encode in FHE).
• The procedure consists a Ring-to-Field conversion, a FHE SIMD based element-wise product, a Field-to-Ring conversion (with or without free truncation).
• set flag0 to 0 : element-wise product of a secret share vector with a private vector (hold by Server)
• set flag0 to 1 : element-wise product of two secret share vectors
• set flag1 to 0 : preform exactly field-to-ring convertion
• set flag1 to 1 : perform free truncate while field-to-ring convertion

Mimic an WAN setting within LAN on Linux

• To use the throttle script under scripts/throttle.sh to limit the network speed and ping latency (require sudo)
• For example, run sudo scripts/throttle.sh wan on a Linux OS which will limit the local-loop interface to about 1Gbps bandwidth and 40ms ping latency. You can check the ping latency by just ping 127.0.0.1. The bandwidth can be check using extra iperf command.

Citing

@ARTICLE{11297783,
author={Dong, Ye and Lu, Wen-jie and Hou, Xiaoyang and Yang, Kang and Liu, Jian},
journal={ IEEE Transactions on Dependable and Secure Computing },
title={{ M&M: Secure Two-Party Machine Learning through Modulus Conversion and Mixed-Mode Protocols }},
year={5555},
volume={},
number={01},
ISSN={1941-0018},
pages={1-15},
doi={10.1109/TDSC.2025.3643323},
url = {https://doi.ieeecomputersociety.org/10.1109/TDSC.2025.3643323},
publisher={IEEE Computer Society},
address={Los Alamitos, CA, USA},
month=dec}