Generative Adversarial Networks (GANs) have been widely applied in different scenarios thanks to the development of deep neural networks. The original GAN was proposed based on the non-parametric assumption of the infinite capacity of networks. However, it is still unknown whether GANs can fit the target distribution without any prior information. Due to the overconfident assumption, many issues remain unaddressed in GANs training, such as non-convergence, mode collapses, and gradient vanishing. Regularization and normalization are common methods of introducing prior information to stabilize training and improve discrimination. Although a handful number of regularization and normalization methods have been proposed for GANs, to the best of our knowledge, there exists no comprehensive survey that primarily focuses on objectives and development of these methods, apart from some incomprehensive and limited-scope studies. In this work, we conduct a comprehensive survey on the regularization and normalization techniques from different perspectives of GANs training. First, we systematically describe different perspectives of GANs training and thus obtain the different objectives of regularization and normalization. Based on these objectives, we propose a new taxonomy. Furthermore, we compare the performance of the mainstream methods on different datasets and investigate the applications of regularization and normalization techniques that have been frequently employed in state-of-the-art GANs. Finally, we highlight potential future directions of research in this domain. Code and studies related to the regularization and normalization of GANs in this work are summarized at https://github.com/iceli1007/GANs-Regularization-Review .

Generative Adversarial Networks (GANs) are the most popular image generation models that have achieved remarkable progress on various computer vision tasks. However, training instability is still one of the open problems for all GAN-based algorithms. Quite a number of methods have been proposed to stabilize the training of GANs, the focuses of which were respectively put on the loss functions, regularization and normalization technologies, training algorithms, and model architectures. Different from the above methods, in this paper, a new perspective on stabilizing GANs training is presented. It is found that sometimes the images produced by the generator act like adversarial examples of the discriminator during the training process, which may be part of the reason causing the unstable training of GANs. With this finding, we propose the Direct Adversarial Training (DAT) method to stabilize the training process of GANs. Furthermore, we prove that the DAT method is able to minimize the Lipschitz constant of the discriminator adaptively. The advanced performance of DAT is verified on multiple loss functions, network architectures, hyper-parameters, and datasets. Specifically, DAT achieves significant improvements of 11.5% FID on CIFAR-100 unconditional generation based on SSGAN, 10.5% FID on STL-10 unconditional generation based on SSGAN, and 13.2% FID on LSUN-Bedroom unconditional generation based on SSGAN.¹

It has become common to collect data from the Internet or hire a third party to train Deep Neural Networks (DNNs). Unfortunately, recent studies have shown that these operations provide a viable pathway for injecting hidden backdoors into DNNs. Several defense methods have been developed to detect malicious samples, with the common assumption that the latent representations of benign and malicious samples exhibit different distributions. However, it is still an open question whether this assumption holds up. In this paper, we investigate such differences via answering three questions: 1) What are the characteristics of the distributional differences 2) How can they be effectively reduced 3) What impact does this reduction have on difference-based defense methods First, the distributional differences of multi-level representations on the regularly trained backdoored models are verified to be significant by adopting Maximum Mean Discrepancy (MMD), Energy Distance (ED), and Sliced Wasserstein Distance (SWD) as the metrics. Then, ML-MMDR, a difference reduction method that adds multi-level MMD regularization into the loss, is proposed, and its effectiveness is testified on three typical difference-based defense methods. The F1 scores of these methods drop from 90%-100% on the regularly trained backdoored models to 60%-70% on the models trained with ML-MMDR.

Generative Adversarial Networks (GANs) have the ability to generate images that are visually indistinguishable from real images. However, recent studies have revealed that generated and real images share significant differences in the frequency domain. In this article, we argue that the frequency gap is caused by the high-frequency sensitivity of the discriminator. According to our observation, during the training of most GANs, severe high-frequency differences make the discriminator focus on high-frequency components excessively, which hinders the generator from fitting the low-frequency components that are important for learning images’ content. Then, we propose two simple yet effective image pre-processing operations in the frequency domain for eliminating the side effects caused by high-frequency differences in GANs training: High-frequency Confusion (HFC) and High-frequency Filter (HFF). The proposed operations are general and can be applied to most existing GANs at a fraction of the cost. The advanced performance of the proposed operations is verified on multiple loss functions, network architectures, and datasets. Specifically, the proposed HFF achieves significant improvements of 42.5% FID on CelebA (128*128) unconditional generation based on SNGAN, 30.2% FID on CelebA unconditional generation based on SSGAN, and 69.3% FID on CelebA unconditional generation based on InfoMAXGAN. Furthermore, we also adopt HFF as the first attempt at data augmentation in the frequency domain for contrastive learning, achieving state-of-the-art performance on unconditional generation. Code is available athttps://github.com/iceli1007/HFC-and-HFF.