Asynchronous Federated Learning with Differential Privacy for Edge Intelligence

Multi-stage adjustable private algorithm(MAPA) by computational differential privacy (DP) to protect the training process,focusing on the impacts of noise on the machine learning in AFL and a better trade-off between the model accuracy and privacy guarantee to improve the trade-off by dynamically adjusting the variance of noise

In order to compare the algorithm, we set up three comparison methods, asgd、audp and mapa. And the algorithm is encapsulated in the docker container and defined by the makefile.

1.Environment Deployment

OS	Ubuntu 18.04
Docker	18.09.7
Docker-compose	1.24.1
OpenSSH	7.6

2. Project structure

├──cloud
|    ├── cloud-asgd.py       
|    ├── cloud-mapa.py            
|    ├── cloud-audp.py              
|    ├── Dockerfile               Docker image build file
|    ├── docker-compose.yml       Defining YAML files for services, networks, and volumes
|    ├── params.py
|    └── sources.list
|
├──edge
|    ├── Dockerfile    
|    ├── docker-compose.yml   
|    ├── sources.list
|    ├── edge-asgd.py
|    ├── edge-mapa.py
|    ├── edge-audp.py
|    ├── data                               Datasets
|    ├   └──MNIST
|    ├      └──  …
|    └──  result                              Output
|        └── [EDGE_NUM:3][METHOD:asgd]        Naming rules: number of edge nodes, number of containers
├──result
├──Makefile                          Setting parameters    
├──README.md
└──ssh_config

3.Network Configuration

Before the application runs, first perform the network configuration of the user operating device and the working node, and write the network topology to the ssh_config file, such as：

Host cloud
    HostName xxx.xxx.xxx.xxx
    Port 22
    User node2user

Host edge1                                     #Node name
    HostName xxx.xxx.xxx.xxx                   #ip
    Port 22                                    #port
    User node1user                             #ssh connection username

4.Parameter settings

Algorithm-related parameters are defined in the Makefile，such as：

# In cloud 
CLOUD := cloud                
MQTT_IP ?= xxx.xxx.xxx.xxx        #cloud device ip
MQTT_PORT ?= 1884               #cloud device MQTT broker port

# In edge  
EDGES := edge1                 #edge device name
EDGES_NUM ?=1                   #edge device numbers
CONTAINER_NUM ?=3             #Number of containers per edge device
METHOD ?= asgd                #asgd,mapa,audp
BATCH_SIZE ?= 24
EPOCH ?= 1
TEST_NUM ?= 100               #Test every TEST_NUM iterations
DATA_ROOT ?= './data/'            #Dataset directory 
RESULT_ROOT ?= './result/'             #result directory

5.Run

Under the project folder, execute the shell command to apply the project:

Perform network configuration between the user-operated device and the working node, including the configuration of ssh-free operation. After completing this step, the user operates the device to perform algorithm-related steps on other working nodes directly through ssh.

make net_config

The program and algorithm data packets are transmitted from the user operating device to the corresponding working node and mirrored.

make build

Start the container, transfer the parameters to each working node, assign tasks to the working nodes, and start the algorithm work.

make run

Look at the work log on the master node, and the log content is dynamically updated in real time following the current program execution.

make logs

Output the experimental results to the host.

make result

Clean up all containers and free up resources.

make clean

niklausliu / Mapa