PrivateSNN

Pytorch Implementation for [PrivateSNN: Privacy-Preserving Spiking Neural Networks].

Accepted in AAAI2022.

Prerequisites

• Python 3.8
• PyTorch 1.5.0
• NVIDIA GPU (>= 12GB)

Getting Started

Training and testing

[STEP 1] Training ANN (ref: https://github.com/nitin-rathi/hybrid-snn-conversion)

Trainig ANN architecture with Backpropagation.

python ann.py --dataset 'CIFAR10' -a 'VGG16' --optimizer 'SGD' --batch_size 256 -lr 1e-2 --lr_interval '0.40 0.60 0.80'

[STEP 2] Generate synthetic training samples

Generating synthetic images from pre-trained ANN model.

python ann_image_gen.py --dataset 'CIFAR10' --num_synimage 1000

[STEP 3] Datafree Conversion

Converting pre-trained ANN model to SNN model with synthetic dataset (from STEP2).

python datafree_conversion.py -a 'VGG16' --dataset 'CIFAR10' --timesteps 100 --batch_size 1024 

Citation

Please consider citing our paper:

@article{kim2021privatesnn,
 title={Privatesnn: Privacy-preserving spiking neural networks},
 author={Kim, Youngeun and Venkatesha, Yeshwanth and Panda, Priyadarshini},
 journal={arXiv preprint arXiv:2104.03414},
 year={2021}
}