CS3511 Solo Project: Federated Learning Solution

1 Objectives
• Learn the basics and processes of Federated Learning.
• Simulate and implement the interactions between clients and the cloud server in Federated Learning.
2 Project Statement
The experimental task is to simulate a simple MINIST classification in horizontal Federated Learning mode.
The whole project is divided into the following three steps to help you improve the functionality and get closer to how the federated learning framework works in the real world.
2.1 Stage 1
In federated learning, we suppose that one server and N clients train a classification task collaboratively, which updates model parameters by a) direct data access or b) “.pth” file reading and writing. In stage 1, all clients participate in each round of updates, shown in Figure 1.

Figure 1: Flowchart of Stage 1
2.2 Stage 2
Based on stage 1, let M out of N clients be selected to participate in the update in each round, which is called the partial participation mode.
1. Please randomly select M clients to participate in the update in each round.
2. The program should support different M.
2.3 Stage 3
In stage 1, the server and clients interact through the “.pth” file reading and writing. Here, the server and clients interact using Socket communication. When the server is running, clients can connect to the server. The communication between the client and server should be implemented using Internet-domain socket.
Specifically, the tasks of the server and each client are as follows: The server:
• Listen on the port, accept local models from each client, and gradually complete model aggregation.
• Send the global model to each client after aggregation.
• Using TCP protocol is suggested.
The client:
• Train locally on its own private dataset and send local model parameters to the server.
• Receive global model parameters from the server.
3 Implementation Details
• The clients’ local private training data will be given together later, with the file named as Client1, Client2, ···, Client20 (there are 20 clients, N = 20). In Stage 1, all 20 clients will participate in the model parameter update. In Stage 2, M of the 20 clients will participate in the model parameter update. In addition, we will provide a Test dataset for the aggregated global model performance evaluation. • Print out as much information as possible to show the intermediate results of your programs.
• Python packages you might need for Stage 3: import socket
You can go online and search for “python socket file transfers” for reference.
4 Material to be submitted
• Submit a PDF experiment report to describe design ideas, experimental procedures, running results, reference materials, etc. Using Latex is suggested.
• Compress the source code of the programs into the Prj+StudentID.tar file. Use meaningful names for the file so that the contents of the file are obvious. Enclose a README file that lists the files you have submitted along with an explanation and show that how your program works. Call it PrjREADME.
• Send your Prj+StudentID.tar file to the Canvas.