Conduct data reconstruction attack on partially encrypted updates using MaskCrypt. The attack implementation is borrowed from https://github.com/JonasGeiping/invertinggradients.
Step 1. Install Pytorch
pip3 install torch torchvision
Step 2. Run the attack script with CIFAR10
python inversion_mask.py --model_arch ResNet18 --dataset CIFAR10 --mask_policy [gradient | random] --mask_ratio [0.0~1.0]
where --mask_policy gradient means the encryption mask will be selected by MaskCrypt's gradient-guided method, and --mask_policy random uses random mask. --mask_ratio can be chosen between 0.0 and 1.0.
For MNIST, we need to change the model architecture and dataset.
python inversion_mask.py --model_arch LeNet5 --dataset MNIST --mask_policy gradient --mask_ratio 0.01
The ground truth image will be stored by name ground_truth_$dataset.png, and the reconstructed image will be stored by name $dataset_$mask_policy_$mask_ratio.png