zk-ml / ZEN

zero knowledge proof for NN inference

ZEN: Efficient Zero-Knowledge Proof for Neural Networks

Rust version

We test the code using rustc 1.47.0. Use rustup override set 1.47.0 to specify the rust version for compilation. For the newer version implementation of ZEN in zen-arkworks, please use rustup override set 1.51.0 for cargo build.

Prepare Data

• Under directory ZEN/zk-ml-private-model-baseline/, run mkdir test-data and cargo run --bin gen_data to generate the mock inputs for baseline and microbenchmark purposes only. For optimization level 3, we load real quantization parameters generated from ZEN/numpyInferenceEngine/XXNet/. For example, cd to ZEN/numpyInferenceEngine/LeNet_CIFAR10 and run python3.8 LeNet_end_to_end_quant.py --model LeNet_Small. The generated quantized parameters are located at ZEN/numpyInferenceEngine/LeNet_CIFAR10/LeNet_CIFAR_pretrained/. For easily reproducing the results, we have saved a copy of parameters for all combinations of model and dataset in director ZEN/zk-ml-private-model/pretrained_model/

Commitments

REMINDER: before benchmarking, please use the corresponding num_window parameter. For different input image size, the commitment setting is different. Refer to pedersen_commit.rs and please ensure you use the correct num_window parameter for corresponding dataset.

• Pedersen input size = window_size * num_window / 8
• pub const PERDERSON_WINDOW_SIZE: usize = 25; // this is for 28X28X1 MNIST input
• pub const PERDERSON_WINDOW_SIZE: usize = 100; // this is for 32X32X3 CIFAR u8 input
• pub const PERDERSON_WINDOW_SIZE: usize = 100; // this is for 46X56X1 FACE u8 input

Microbenchmarks(under zk-ml-baseline directory)

Conv and FC layers different levels of optimization

• cargo run --bin microbench_conv_layered_optimization_by_kernel_size --release 2>/dev/null
• cargo run --bin microbench_fc_layered_optimization --release 2>/dev/null

SIMD (stranded encoding) under different batch size

• cargo run --bin microbench_SIMD_by_batch_size --release 2>/dev/null

LeNet Small on CIFAR dataset different levels of optimization

• cargo run --bin microbench_lenet_small_cifar_naive --release 2>/dev/null
• cargo run --bin microbench_lenet_small_cifar_op1 --release 2>/dev/null
• cargo run --bin microbench_lenet_small_cifar_op2 --release 2>/dev/null
• cargo run --bin microbench_lenet_small_cifar_op3 --release 2>/dev/null

Naive/baseline for all combinations of models and datasets(only calculate the number of constraints)

• cargo run --bin shallownet_naive_mnist --release 2>/dev/null
• cargo run --bin lenet_small_naive_pedersen_cifar --release 2>/dev/null
• cargo run --bin lenet_medium_naive_pedersen_cifar --release 2>/dev/null
• cargo run --bin lenet_small_naive_pedersen_face --release 2>/dev/null
• cargo run --bin lenet_medium_naive_pedersen_face --release 2>/dev/null
• cargo run --bin lenet_large_naive_pedersen_face --release 2>/dev/null

Under zk-ml-private-model directory

Optmization level 3 for all combinations of models and datasets

• cargo run --bin shallownet_optimized_pedersen_mnist --release 2>/dev/null
• cargo run --bin lenet_small_optimized_pedersen_cifar --release 2>/dev/null
• cargo run --bin lenet_medium_optimized_pedersen_cifar --release 2>/dev/null
• cargo run --bin lenet_small_optimized_pedersen_face --release 2>/dev/null
• cargo run --bin lenet_medium_optimized_pedersen_face --release 2>/dev/null
• cargo run --bin lenet_large_optimized_pedersen_face --release 2>/dev/null

Under zk-ml-accuracy directory

Optmization level 3 for all combinations of models and datasets

• cargo run --bin shallownet_accuracy --release 2>/dev/null
• cargo run --bin lenet_small_cifar_accuracy --release 2>/dev/null
• cargo run --bin lenet_medium_cifar_accuracy --release 2>/dev/null
• cargo run --bin lenet_small_face_accuracy --release 2>/dev/null
• cargo run --bin lenet_medium_face_accuracy --release 2>/dev/null
• cargo run --bin lenet_large_face_accuracy --release 2>/dev/null