IFL-GAN: Improved Federated Learning Generative Adversarial Network With Maximum Mean Discrepancy Model Aggregation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) typically require centralized, independent and identically distributed (i.i.d.) data.
• Real-world data is often distributed across multiple clients, posing challenges for data aggregation due to bandwidth, privacy, or storage constraints.
• Existing federated learning GAN (FL-GAN) approaches struggle with non-i.i.d. data, leading to convergence difficulties and low-quality generated samples.

Purpose of the Study:

• To propose a novel approach for training GANs on distributed, non-i.i.d. data.
• To overcome the limitations of federated averaging in FL-GAN by introducing a new aggregation strategy.
• To improve the convergence speed and quality of generated instances in distributed GAN training.

Main Methods:

• Developed an improved FL-GAN (IFL-GAN) that aggregates locally trained generators' updates using Maximum Mean Discrepancy (MMD).
• This MMD-based aggregation allows local GANs to have different weights, facilitating faster global model convergence.
• Evaluated IFL-GAN on benchmark datasets including MNIST, CIFAR10, and SVHN.

Main Results:

• IFL-GAN demonstrated significantly faster convergence compared to traditional federated averaging methods.
• The proposed method achieved superior performance in generating high-quality instances.
• Experimental results showed the highest inception scores among tested approaches.

Conclusions:

• IFL-GAN effectively addresses the challenges of training GANs on non-i.i.d. distributed data.
• The MMD-based aggregation strategy offers a robust alternative to federated averaging for distributed GAN training.
• The approach shows significant promise for real-world applications requiring privacy-preserving and efficient distributed GAN model development.