PerFED-GAN: Personalized Federated Learning via Generative Adversarial Networks

Xingjian Cao, Gang Sun, Hongfang Yu, Mohsen Guizani

Abstract

Federated learning is gaining popularity as a distributed machine learning method that can be used to deploy AI-dependent IoT applications while protecting client data privacy and security. Due to the differences of clients, a single global model may not perform well on all clients, so the personalized federated learning method, which trains a personalized model for each client that better suits its individual needs, becomes a research hotspot. Most personalized federated learning research, however, focuses on data heterogeneity while ignoring the need for model architecture heterogeneity. Most existing federated learning methods uniformly set the model architecture of all clients participating in federated learning, which is inconvenient for each client's individual model and local data distribution requirements, and also increases the risk of client model leakage. This paper proposes a federated learning method based on co-training and generative adversarial networks (GANs) that allows each client to design its own model to participate in federated learning training independently without sharing any model architecture or parameter information with other clients or a center. In our experiments, the proposed method outperforms the existing methods in mean test accuracy by 42% when the client's model architecture and data distribution vary significantly.

Introduction

Standard Federated Learning (FL) typically requires every participant (client) to use the exact same neural network architecture to facilitate weight aggregation (e.g., FedAvg). However, this "one-size-fits-all" approach faces significant hurdles in real-world deployments. First, Data Heterogeneity (Non-IID) means a single global model often fails to generalize across diverse local data distributions. Second, Model Heterogeneity arises because different organizations may have varying hardware constraints or proprietary model designs they wish to keep private.

Existing personalized FL methods often still require consistent architectures or rely on sharing model parameters, which can leak intellectual property or sensitive data. PerFED-GAN addresses these challenges by allowing clients to remain "black boxes," maintaining their specific architectures and parameters entirely locally.

Methodology: The PerFED-GAN Framework

PerFED-GAN utilizes a strategy based on co-training and Generative Adversarial Networks (GANs). Instead of aggregating gradients or weights at a central server, the framework uses synthetic data as a medium for knowledge distillation.

The PerFED-GAN architecture showing the interaction between multiple heterogeneous clients and the central server.

The process consists of four primary stages:

  1. Initial Local Training: Each client $i$ trains their unique model $M_i$ on their private dataset $Tr_i$.
  2. GAN Training: The client uses their trained model as a discriminator ($D$). They train a local generator ($G$) to produce synthetic data samples that "fool" the model into classifying them into target categories.
  3. Communication and Aggregation: Clients generate a synthetic dataset $A_i$ and upload these samples to a central server. The server merges these into a diverse "knowledge pool" $Ca$ and redistributes random subsets back to the clients.
  4. In-process Training: Clients update their personalized models using a union of their original private data and the aggregated synthetic data received from the server.
Algorithm 1: The formal procedure for PerFED-GAN training across N parallel clients.

This approach ensures that a tiny mobile model can learn from a massive server-grade model because the medium of exchange is data samples rather than model weights.

Experimental Evaluation

The researchers evaluated PerFED-GAN using three standard computer vision benchmarks: CIFAR-10, CIFAR-100, and FEMNIST. To simulate model heterogeneity, clients were assigned randomly generated CNNs with varying depths and filter counts.

Key Results

  • Accuracy Gains: In Non-IID settings, PerFED-GAN outperformed local training by an average of 35% on CIFAR-10 and 49% on CIFAR-100.
  • Comparison with Baselines: The method outperformed existing personalized FL methods (like FedAvg with fine-tuning and FedProx) by approximately 42% in mean test accuracy when both model architectures and data distributions varied significantly.
  • Privacy and Security: By sharing synthetic samples instead of model weights, the framework mitigates "white-box" attacks. Furthermore, the paper demonstrates that PerFED-GAN can be integrated with Differential Privacy (DP-CGAN) to further protect against data leakage.

Discussion and Conclusion

PerFED-GAN effectively breaks the architecture barrier in federated learning. Its performance is notably stronger when client data is highly diverse, as the synthetic data pool becomes a richer source of global knowledge.

While the method offers significant advantages in flexibility and privacy, it introduces higher local computational overhead due to GAN training. Future work will focus on optimizing the efficiency of local GAN training to better suit low-power edge devices and reducing the communication bandwidth required for transferring synthetic data.

Summary updated

Abstract

Federated learning is gaining popularity as a distributed machine learning method that can be used to deploy AI-dependent IoT applications while protecting client data privacy and security. Due to the differences of clients, a single global model may not perform well on all clients, so the personalized federated learning method, which trains a personalized model for each client that better suits its individual needs, becomes a research hotspot. Most personalized federated learning research, however, focuses on data heterogeneity while ignoring the need for model architecture heterogeneity. Most existing federated learning methods uniformly set the model architecture of all clients participating in federated learning, which is inconvenient for each client's individual model and local data distribution requirements, and also increases the risk of client model leakage. This paper proposes a federated learning method based on co-training and generative adversarial networks(GANs) that allows each client to design its own model to participate in federated learning training independently without sharing any model architecture or parameter information with other clients or a center. In our experiments, the proposed method outperforms the existing methods in mean test accuracy by 42% when the client's model architecture and data distribution vary significantly.
BibTeX
@article{cao2023perfed,
  author = {Xingjian Cao and Gang Sun and Hongfang Yu and Mohsen Guizani},
  title = {{PerFED-GAN: Personalized Federated Learning via Generative Adversarial Networks}},
  journal = {IEEE Internet of Things Journal},
  year = {2023},
  url = {https://openreview.net/forum?id=8swbZu0gX7},
}
Figure