Smaug

Modular Augmentation of LLVM for Secure Multi-Party Computation

Smaug is an LLVM-based compiler that transforms standard C/C++ and Rust programs into secure multi-party computation (MPC) protocols. It automatically converts non-oblivious programs into their oblivious counterparts and applies circuit-level optimizations as LLVM code transformations.

Features

• Language support: Compile from C, C++, and Rust via standard LLVM frontends
• Automatic oblivious transformation: Convert programs with private control flow to trace-oblivious form without programmer annotations
• Circuit optimization: Loop vectorization, loop flattening, loop splitting, and reduction parallelization to minimize circuit depth
• Multi-protocol backends: Supports Yao's garbled circuits and GMW protocol via the EMP toolkit
• LLVM integration: Leverages LLVM's optimization pipeline, error messaging, and tooling

Prerequisites

• Ubuntu 22.04+ (tested on 24.04)
• LLVM 18 (apt install llvm-18 clang-18)
• CMake 3.22+
• OpenSSL development libraries
• EMP toolkit: emp-tool, emp-ot, emp-sh2pc, emp-aby

See docs/install.md for detailed installation instructions.

Building

# Build the MPC runtime
cd runtime && mkdir -p build && cd build
cmake .. && make -j$(nproc)
sudo make install
cd ../..

# Build the LLVM pass plugin
cd passes && mkdir -p build && cd build
cmake .. && make -j$(nproc)
cd ../..

Running Benchmarks

cd benchmarks

# Garbled circuit (Yao) backend
make run

# GMW backend
make run-no-gc

# With loop flattening optimization
make run-loop-flatten
make run-loop-flatten-no-gc

Benchmarks

Program	Description	Input Size
biometric	Biometric matching	N=4096, D=4
convex_hull	Convex hull computation	N=256
count10/count102	Pattern counting	N=4096
db_variance	Database variance	N=256
histogram	Histogram computation	N=4096, D=5
inner_product	Inner product	N=4096
kmeans_iteration	K-means clustering	N=256, k=16
longest102	Longest pattern	N=4096
matmul	Matrix multiplication	N=16
max_dist_between_syms	Maximum distance	N=4096
minimal_points	Minimal points	N=256
mnistRelu	MNIST with ReLU	N=16
psi	Private set intersection	N=1024

Compiling Your Own Programs

See docs/usage.md for a step-by-step guide on compiling MPC programs with Smaug.

Project Structure

smaug/
  passes/       LLVM transformation passes (libpasses.so)
  runtime/      MPC runtime libraries (libgc-opt, libmpc)
  benchmarks/   Benchmark programs and test pipeline
  docs/         Documentation

Citation

If you use Smaug in your research, please cite:

@INPROCEEDINGS{11023285,
  author={Garg, Radhika and Wang, Xiao},
  booktitle={2025 IEEE Symposium on Security and Privacy (SP)}, 
  title={Smaug: Modular Augmentation of LLVM for MPC}, 
  year={2025}}

License

MIT License. See LICENSE.