1: \begin{abstract}

2: \label{sec:abstract}

3: 

4: Generative adversarial networks (GANs) typically require ample data for training in order to synthesize high-fidelity images.

5: %

6: Recent studies have shown that training GANs with limited data remains formidable due to discriminator overfitting, the underlying cause that impedes the generator's convergence.

7: %

8: This paper introduces a novel strategy called Adaptive Pseudo Augmentation (APA) to encourage healthy competition between the generator and the discriminator.

9: %

10: As an alternative method to existing approaches that rely on standard data augmentations or model regularization,

11: APA alleviates overfitting by employing the generator itself to augment the real data distribution with generated images, which deceives the discriminator adaptively.

12: %

13: Extensive experiments demonstrate the effectiveness of APA in improving synthesis quality in the low-data regime. 

14: %

15: We provide a theoretical analysis to examine the convergence and rationality of our new training strategy.

16: %

17: APA is simple and effective. It can be added seamlessly to powerful contemporary GANs, such as StyleGAN2, with negligible computational cost. Code: \url{https://github.com/EndlessSora/DeceiveD}.

18: 

19: \end{abstract}

20: