Patch-SplitNN - Official PyTorch Implementation

Patch-based Privacy Preserving Neural Network for Vision Tasks
This paper is presented on IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2023, and available in WACV2023 open access.

Requirements

• h5py: 3.7.0
• hydra-core: 1.2.0
• lightning-bolts: 0.4.0
• pytorch-lightning: 1.6.1
• torch: 1.12.0
• torchvision: 0.13.0 (>= ver 0.12.0 supports PCAM dataset )
• thop: 0.1.1

For details, see requirements.txt.

Dataset

• CIFAR-10 and CIFAR-100
• PCAM (Patch Camelyon)

Training

On CIFAR

• Move to cifar folder.

$ cd cifar

• ResNet (Reference model) training.

$ python resnet_cifar_main.py

• Patch SplitNN

$ python patch_splitnn_cifar_main.py adapt_net=False

• Patch SplitNN+

$ python patch_splitnn_cifar_main.py adapt_net=True

On PCAM

• Before training, dataset should be downloaded. Then set the path to YAML.

• Move to pcam folder.

$ cd pcam

• ResNet (Reference model) training.

$ python resnet_pcam_main.py

• Patch SplitNN

$ python patch_splitnn_main.py adapt_net=False

• Patch SplitNN+

$ python patch_splitnn_main.py adapt_net=True

Parameters of YAML

Configuration files are stored in each config/.

• data_dir: path to dataset directory.
• logger: tensorboard or wandb can be used.
• dataset: CIFAR10, CIFAR100, PatchCIFAR10, PatchCIFAR100, or PCAM.
• base_model: resnet18 or resnet34
• patch_size: the size of patch images.
• patch_stride: the stride size for splitting.
• num_uppmodels: the number of upper models for Patch SplitNN.
• adapt_net: if True, Adaptation Net is set.
• upp_loss_ratio: Coefficient value for loss function of upper loss.
• drop_rate: the ratio of dropping patches.

Docker

NVIDIA Container Toolkit is needed.

• Build docker image

$ docker build -f docker/Docker . -t <image_name>

• Run docker image

$ docker run --rm \
             --gpus all \
             -v /path/to/patch-splitnn:/opt/patch-splitnn \
             -it <image_name> \
             /bin/bash
$ cd /opt/patch-splitnn

Citation

@InProceedings{Mabuchi_2023_WACV,
    author    = {Mabuchi, Mitsuhiro and Ishikawa, Tetsuya},
    title     = {Patch-Based Privacy Preserving Neural Network for Vision Tasks},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {1550-1559}
}