FedSDP: Federated Self-Derived Prototypes for Personalized Federated Learning

📌 Introduction

Federated Learning (FL) is a privacy-preserving machine learning paradigm that enables multiple clients to train collaboratively without sharing raw data. However, traditional FL models struggle with non-IID data distributions, leading to degraded performance.

FedSDP (Federated Self-Derived Prototypes) is a Personalized Federated Learning (PFL) framework designed to dynamically balance generalization and personalization through:

1. Global-Local Similarity Weight (GL-Sim Weight): Adjusts personalization level based on global-local feature similarity.
2. Personalization Early Stopping Indicator (P-Stop Indicator): Prevents overfitting by dynamically stopping personalization at the optimal point.

Our experiments show that FedSDP outperforms state-of-the-art PFL frameworks across multiple datasets in non-IID settings.

🔧 Training

To train FedSDP on CIFAR-10 with 20 clients, use the following command:

python ./main.py -data cifar10 -nb 10 -m resnetfc2 -lbs 8 -lr 0.01 -optim sgd -gr 100 -algo FedSDP -jr 1 -nc 20 -dir ./dataset/cifar10/dir01 -lda 0.35 -did 0 -ls1 1 -ls2 1 -ls3 1

📌 Explanation of Arguments

Argument	Description
-data cifar10	Dataset used (CIFAR-10)
-nb 10	Number of classes
-m resnetfc2	Model architecture (ResNet with bridge layers)
-lbs 8	Local batch size (8)
-lr 0.01	Learning rate (0.01)
-optim sgd	Optimizer used (SGD)
-gr 100	Global rounds (100)
-algo FedSDP	Algorithm used (FedSDP)
-jr 1	Joint training ratio (1)
-nc 20	Number of clients (20)
-dir ./dataset/cifar10/dir01	Dataset directory
-lda 0.35	Scaling factor for GL-Sim Weight
-did 0	Device ID (GPU/CPU selection)
-ls1 1 -ls2 1 -ls3 1	Local Epoch for body, bridge, and head