Shark

Shark is a framework for actively secure inference of machine learning models. This repository contains implementation of protocols from our IEEE S&P 2025 paper.

Building

Shark requires CMake (>= 3.16) and a C++ compiler (with support for C++20 standard). Shark also requires Eigen3 which can be installed using sudo apt install libeigen3-dev (on Ubuntu/Debian) or brew install eigen (on macOS/linuxbrew). Shark optionally requires OpenMP for parallelization.

To compile Shark, run the following command from the project root:

cmake -DCMAKE_BUILD_TYPE=Release -S . -B build/
cmake --build build/ --config Release --target all -j

Benchmarking

We provide four benchmarking scripts in the root of project that can reproduce the benchmarking results from the paper (specifically, Table 2 and 3).

Local Setup

In case you want to run both parties of Shark on a single machine, with each party using 4 threads each, the following commands can be used to run benchmarks and microbenchmarks, respectively.

bash run-benchmarks-local.sh
bash run-microbenchmarks-local.sh

Two Machine Setup

If you have two machines connected through IP, lets say at 10.0.0.1 and 10.0.0.2, run the following commands on each machine to run benchmarks:

bash run-benchmarks-remote.sh 0 10.0.0.2 # on 10.0.0.1
bash run-benchmarks-remote.sh 1 10.0.0.1 # on 10.0.0.2

Microbenchmarks can be run by replacing script with run-microbenchmarks-remote.sh. By default, Shark assumes the port range 42000-42100 to be open. In case you would like to use a different range, provide the starting port as the third argument to the scripts.

Custom Model Benchmarking

To write a custom model in our framework, please refer to benchmarks/mnist-A.cpp for reference. To build your custom model model.cpp in Shark, place it in benchmarks/ directory, add the following to the end of CMakeLists.txt:

add_executable(benchmark-model benchmarks/model.cpp)
target_link_libraries(benchmark-model ${PROJECT_NAME})

Rebuild Shark and your binary would be available as build/benchmark-model. To run it, use the following commands:

./build/benchmark-model 2 # on both machines, to generate keys
OMP_NUM_THREADS=4 ./build/benchmark-model 0 10.0.0.2 # on 10.0.0.1
OMP_NUM_THREADS=4 ./build/benchmark-model 1 10.0.0.1 # on 10.0.0.2

Bugs

Please feel free to open an issue in case you encounter a bug.

Citation

If you make use of our work, please consider citing us:

@INPROCEEDINGS {
    author = { Gupta, Kanav and Chandran, Nishanth and Gupta, Divya and Katz, Jonathan and Sharma, Rahul },
    booktitle = { 2025 IEEE Symposium on Security and Privacy (SP) },
    title = {{ Shark: Actively Secure Inference using Function Secret Sharing }},
    year = {2025},
    doi = {10.1109/SP61157.2025.00175},
}