MARLUT

This repo implements Marlut, a maliciously secure lookup table (LUT) protocol over $\mathbb{Z}_{2^k}$ or $\mathbb{GF}(2^k)$.

We regret that the code might not be the most concise or readable as it contains tons of SIMD optimizations and employs metaprogramming techniques.

Setup and Building

Rust Nightly is required. Evaluations in our paper were run with rustc 1.91.0-nightly (040a98af7 2025-08-20).

Our protocols support 2- and 3-dimensional benchmarks. You can build them with:

# 2-dimensional
cargo build --release
# 3-dimensional
cargo build --release --no-default-features --features "clmul GF8_3 GF8NonPre_3 FP16Exp_3 FP8Add_3 Boolean_3"

AVX512 is optional, but machines with AVX512 support will exhibit significantly better performance due to tailored optimizations.

Benchmarking

The CLI for the benchmark binary (run target/release/marlut -h) offers some description and help on the parameters. We have one single protocol with two implementations:

• lut_sp / lut_sp_malsec: Generic protocol for $\mathbb{Z}_{2^k}$ and $\mathbb{GF}(2^k)$, but optimized for $\mathbb{Z}_{2^k}$. Malicious security for $\mathbb{Z}_{2^k}$.
• lut_sp_boolean / lut_sp_boolean_malsec: Optimized for $\mathbb{GF}(2^k)$. Malicious security for $\mathbb{GF}(2^k)$.

Benchmarks for our protocol are as follows

• with semi-honest security

  • lut256-sp: Evaluating single-input tables of size $2^8$ over ring $\mathbb{Z}_{2^8}$.
  • lut256-sp-fp16-exp: Evaluating the exponential function with a single-input table of size $2^{16}$ over ring $\mathbb{Z}_{2^{16}}$. The table can be replaced to compute the Sigmoid function.
  • lut256-sp-fp8-add: Evaluating floating-point addition with a two-input table of size $2^{16}$ over $\mathbb{Z}_{2^{8}}$. The table can be replaced to compute floating-point multiplication.
  • lut256-sp-gf8: Evaluating a single-input table of size $2^8$ over $\mathbb{GF}(2^8)$. This uses the generic implementation.
  • lut256-sp-boolean: Evaluating a single-input table of size $2^8$ over $\mathbb{GF}(2^8)$. This uses the optimized boolean implementation.

• with active security

  • mal-lut256-sp: The maliciously secure version of lut256-sp.
  • mal-lut256-sp-fp16-exp: The maliciously secure version of lut256-sp-fp16-exp.
  • mal-lut256-sp-fp8-add: The maliciously secure version of lut256-sp-fp8-add.
  • mal-lut256-sp-boolean: The maliciously secure version of lut256-sp-boolean.

In addition, most benchmarks from MAESTRO are retained. For a fair comparison, it is advised to run our version because we have made some optimizations to the networking framework.

To run the benchmarks, you may refer to run_bench.sh. You should run the script on the three machines

./run_bench.sh p1.toml # Machine 1
./run_bench.sh p2.toml # Machine 2
./run_bench.sh p3.toml # Machine 3

Remember to update the machine IP addresses in p{1,2,3}.toml. For more details, please also refer to README_MAESTRO.md.