The implementation of the paper "Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning using Independent Component Analysis". In order to use the functions implemented, please refer directly to the implementation. For the attack implementation, there will be explanations for each arguement in both cocktail_party_attack/attack.py and cocktail_party_attack/hyperparameter_optimization.py.
By using the below command you can create a conda environment to use this repository
conda env create -f environment.yml
To train a model and save it, one can use train_all function located in cocktail_party_attack/train.py. However, before using it the data loaders should be received.
One can utilize this function by
from cocktail_party_attack import train_allFor now there are functions to retrieve data loaders of datasets MNIST, CIFAR10, CIFAR100, and Tiny ImageNet. Those functions are load_mnist_dataloaders, load_cifar10_dataloaders, load_cifar100_dataloaders, and load_tiny_imagenet_dataloaders, respectively. All the functions are located under cocktail_party_attack/data.py and one can use those functions using
from cocktail_party_attack import load_tiny_imagenet_dataloaders # function name By using CocktailPartyAttack class defined in cocktail_party_attack/cpa.py, you can conduct attack py following the explanations. Also, you can conduct hyperparameter optimization experiments for attack performance depending on either mean PSNR, mean LPIPS or individual (target image) LPIPS metrics.
Again, you can import this class using
from cocktail_party_attack import CocktailPartyAttack