🔥🔥 [2026] News: We now open-source our private inference code at https://github.com/Tianshi-Xu/SEC-PPDL

🔥 PrivCirNet is accepted by NeurIPS'24!!

🔥 Introducing a network-protocol co-optimization framework for accelerating homomorphic encrypted network inference.

PrivCirNet zoo

We provide PrivCirNet models which can be downloaded here.

Installation

1. Install torch>=2.0.0, python>=3.10.
2. Install other pip packages via pip install -r requirements.txt.
3. Prepare the dataset including CIFAR-10, CIFAR100, Tiny ImageNet and ImageNet.

Training Scripts

Run CirILP.py to get block size configuration

In this step, we load the pretrained model and get the layer-wise block sizes configuration. The script is ILP.sh. For example, to configure ViT on CIFAR-10, you can run the following command:

CUDA_VISIBLE_DEVICES=1 python CirILP.py -c configs/datasets/ViT/cifar10_ILP.yml --model vit_7_4_32 PATH_TO_CIFAR10

A log file named vit_c10_ILP.log will be created to save the configuration and logs. You will get the layer-wise block sizes configuration in the log file.

There are two parameters in the .yml file you need to specify:

• budget. You can set the budget to 2, 4, and 8 whose latency is less than that of networks with uniform block size 2, 4, and 8.
• better_initialization. True means use the initialization method proposed in PrivCirNet, False means use the previous initialization method, i.e., $\min |W'-W|_2^2$.

Train the circulant models

In this step, we load the pretrained model and train the circulant models with the layer-wise block sizes configuration obtained from the previous step.

This step is simple which is the same as training the original models. The script is train_cir.sh. For example , to train ViT on CIFAR-10, you can run the following command:

CUDA_VISIBLE_DEVICES=1 python train_cir.py -c configs/datasets/ViT/cifar10_fix.yml --model vit_7_4_32 PATH_TO_CIFAR10

A log file named your_log_name.log will be created to save logs. The checkpoints will be saved in the output folder. There are several parameters in the .yml file you need to specify:

• fix_blocksize_list. It is the block size configuration you get from the previous step. It must be separated by commas.
• log_name. The name of the log file.
• use_kd. Whether to use knowledge distillation. If True, you need to specify the teacher_model and teacher_checkpoint in the .yml file.
• initial_checkpoint. You must load the initial checkpoint from the original model. All initial checkpoints can be downloaded here.

Train other baseline models

This repository also provides the training scripts for the two baselines:

• Uniform block size circulant networks. The training script is the same as Train the circulant models where you can set fix_blocksize to 2, 4, and 8 to train uniform block size circulant networks.
• SpENCNN (structured pruning). The training script is train_prune.sh. For example, to train MobileNetV2 on CIFAR-10, you can run the following command:

CUDA_VISIBLE_DEVICES=1 python train_prune.py -c configs/datasets/Prune/cifar10.yml --model c10_prune_mobilenetv2 PATH_TO_CIFAR10

• You can set prune_ratio in the .yml file to specify the pruning ratio.
• Pruning models can be downloaded here.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If our work assists your research, feel free to give us a star ⭐ or cite us using:

@inproceedings{NEURIPS2024_ca987391,
 author = {Xu, Tianshi and Wu, Lemeng and Wang, Runsheng and Li, Meng},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang},
 pages = {111802--111831},
 publisher = {Curran Associates, Inc.},
 title = {PrivCirNet: Efficient Private Inference via Block Circulant Transformation},
 url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/ca9873918aa72e9033041f76e77b5c15-Paper-Conference.pdf},
 volume = {37},
 year = {2024}
}