Rainysponge / Split-learning-Attacks

Split-learning-Attacks

Tree

SLFrame
│ config.yaml
│ dataTest.py
│ log.txt
│ Test.py # main │
├─core
│ │ splitApi.py
│ │ splitApiO.py
│ │
│ ├─communication
│ │ │ base_com_manager.py
│ │ │ message.py
│ │ │ msg_manager.py
│ │ │ observer.py
│ │ │ init.py
│ │ │
│ │ ├─mpi
│ │ │ │ mpi_com_mananger.py
│ │ │ │ mpi_receive_thread.py
│ │ │ │ mpi_send_thread.py
│ │ │ │ init.py
│ │
│ │
│ ├─dataset …… │ │ │ abstractDataset.py
│ │ │ abstractDatasetFactory.py
│ │ │ baseDataset.py
│ │ │ baseDatasetFactory.py
│ │ │ datasetFactory.py
│ │ │
│ │ ├─controller
│ │ │ │ adultController.py
│ │ │ │ cheXpertController.py
│ │ │ │ cifar10Controller.py
│ │ │ │ fashionmnistController.py
│ │ │ │ germanController.py
│ │ │ │ mnistController.py
│ │ │
│ │ │
│ │ ├─dataset
│ │ │ │ adult.py
│ │ │ │ cheXpert.py
│ │ │ │ cifar10.py
│ │ │ │ fashionmnist.py
│ │ │ │ german.py
│ │ │ │ mnist.py
│ │ │
│ │ ├─partition │ │ │ │ basePartition.py
│ │ │ │ basePartitionFactory.py
│ │ │ │ cifar10Partition.py
│ │ │ │ partitionFactory.py
│ │ │ │ partitionUtils.py
│ │
│ ├─log
│ │ │ abstractLog.py
│ │ │ baseLog.py
│ │ │ Log.py
│ │
│ │
│ ├─model
│ │ │ Alex_model.py
│ │ │ cnn.py
│ │ │ DenseNet.py
│ │ │ EffientNet0.py
│ │ │ modelFactory.py
│ │ │ models.py
│ │ │ models_for_U.py
│ │ │ model_factory.py
│ │ │ resnet.py
│ │
│ ├─process
│ │ baseClient.py
│ │ baseClientManager.py
│ │ baseServer.py
│ │ baseServerManageger.py
│ │
│ ├─variants
│ │ │ variantsFactory.py
│ │ │
│ │ ├─Asynchronous
│ │ │
│ │ ├─asyVanilla │ │ │ │ client.py
│ │ │ │ client_manager.py
│ │ │ │ message_define.py
│ │ │ │ readme.md
│ │ │ │ server.py
│ │ │ │ server_manager.py
│ │ │ │ init.py
│ │ │
│ │ ├─comp_model
│ │ │
│ │ ├─fedavg
│ │ │
│ │ │
│ │ ├─parallel_U_Shape
│ │ │
│ │ ├─SGLR
│ │ │
│ │ ├─SplitFed
│ │ │
│ │ ├─SplitFed2
│ │ │
│ │ ├─synchronization
│ │ │
│ │ ├─TaskAgnostic
│ │ │
│ │ ├─Ushaped
│ │ │
│ │ ├─vanilla
│ │ │
│ │ ├─vertical
│
│
├─data
│ ├─adult
│ │ adult.data
│ │ adult.test
│ │
│ ├─CheXpert-v1.0-small
│ ├─cifar10
│ │
│ ├─fashionmnist
│ │ └─FashionMNIST
│ │
│ ├─german
│ │ german.data-numeric
│ │
│ └─mnist
│ │
│ └─MNIST
│
├─model_save
│
└─Parse
│ abstractParse.py
│ baseParse.py
│ parseFactory.py
│ utlis.py
│
├─JSON
│ │ jsonParse.py
│
│
├─YAML …… YAML │ │ yamlParse.py

How to Use

After downloading the entire file, it is necessary to configure the relevant environment and install MPI along with the following libraries:

yaml
py4mpi
pytorch
scikit-learn
pandas
numpy
pillow
matplotlib

To use existing paradigms, only the config.yaml file needs to be configured. Here's an example using PSL as an illustration:

dataDir: ./data/
dataset: mnist
device: cpu
download: true
batch_size: 128
epochs: 15
log_step: 20
lr: 0.01
max_rank: 2
partition_method: homo
variants_type: psl
server_rank: 0

Among these parameters, variants_type indicates the type of paradigm, partition_method represents the data set partitioning method, dataset denotes the used dataset, and epochs and batch_size respectively specify the number of training rounds and the batch size per round.

An example of the main function is provided below:

if __name__ == '__main__':
    args = parseFactory(fileType=YAML).factory()
    client_model = LeNetClientNetwork()
    
    server_model = LeNetServerNetwork()
    args.load('./config.yaml')

    comm, process_id, worker_number = SplitNN_init(args)
    args["rank"] = process_id  # Set the current process_id.

    args["client_model"] = client_model
    args["server_model"] = server_model
    device = init_training_device(process_id, worker_number - 1, args.gpu_num_per_server)
    args["device"] = device

    dataset = datasetFactory(args).factory()  # loader data and partition method

    train_data_num, train_data_global, test_data_global, local_data_num, \
    train_data_local, test_data_local, class_num = dataset.load_partition_data(process_id) 

    SplitNN_distributed(process_id, args)

In the main function, it is necessary to declare the used model.

    args["client_model"] = client_model
    args["server_model"] = server_model

To execute, input the following command in the command line: mpiexec -np N python filename Here, N represents the number of processes, equivalent to the number of clients plus one, and 'filename' indicates the name of the file where the main function is located.

Your Own Paradigm

You can imitate the writing style of your own paradigms. The required files are as follows:

client.py  // Defines relevant parameters and training methods for the client
client_manager.py  // Defines methods for client communication
message_define.py  // Defines relevant parameters for communication
server.py  // Defines relevant parameters and training methods for the server
server_manager.py  // Defines methods for server communication

You can also inherit relevant classes from the process to use some common methods.