Adversarial symmetric GANs: Bridging adversarial samples and adversarial networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative adversarial networks (GANs) show strong performance but struggle with training instability.
• Existing training strategies have not fully resolved the challenge of GAN training instability.
• Vanilla GANs perform adversarial training on fake samples but overlook training on real samples.

Purpose of the Study:

• To investigate the root causes of GAN training instability from the perspective of adversarial samples.
• To propose a novel GAN architecture, adversarial symmetric GANs (AS-GANs), to improve training stability and performance.
• To provide theoretical explanations for the effectiveness of AS-GANs.

Main Methods:

• Developed adversarial symmetric GANs (AS-GANs) by integrating adversarial training of the discriminator on real samples into the standard GAN framework.
• Empirically validated AS-GANs on various image generation datasets (CIFAR-10, CIFAR-100, CelebA, LSUN) using diverse network architectures.
• Conducted theoretical analysis to elucidate the mechanisms behind AS-GAN's improvements.

Main Results:

• AS-GANs demonstrated significantly stabilized training and accelerated convergence compared to baseline GANs.
• Generated images showed consistent and substantial improvements in Fréchet Inception Distance (FID) scores.
• The discriminator in AS-GANs became more robust, providing more informative gradients with reduced adversarial noise.

Conclusions:

• Incorporating adversarial training on real samples, as done in AS-GANs, effectively addresses GAN training instability.
• AS-GANs offer a robust and efficient approach for stable image generation with improved quality.
• This work bridges adversarial samples and adversarial networks, opening new avenues for future research in generative models.