On the convergence and mode collapse of GAN

Zhang, Zhaoyu; Li, Mengyan; Yu, Jun

doi:10.1145/3283254.3283282

Cited by 34 publications

(14 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…WGAN and WGAN-GP. A common problem in the training of GANs is the so-called mode collapse [15]. When it happens, the generator begins to synthesize the same images regardless of the input data.…”

Section: Gan Architectures Overviewmentioning

confidence: 99%

Comparison of generative adversarial networks architectures forbiomedical images synthesis

Berezsky¹,

Liashchynskyi²

2021

AAIT

View full text Add to dashboard Cite

The article analyzes and compares the architectures of generativeadversarialnetworks. These networks are based on convolu-tional neural networks that are widely used for classification problems. Convolutional networks require a lot of training data to achieve the desired accuracy. Generativeadversarialnetworks are used for the synthesis of biomedical images in this work. Biomedi-cal images are widely used in medicine, especially in oncology. For diagnosis in oncology biomedical images are divided into three classes: cytological, histological, and immunohistochemical. Initial samples of biomedical images are very small. Getting trainingimages is a challenging and expensive process. A cytological training datasetwas used for the experiments. The article considers the most common architectures of generative adversarialnetworks suchas Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN),Wasserstein GAN with gradient penalty (WGAN-GP), Boundary-seeking GAN (BGAN), Boundary equilibrium GAN (BEGAN). A typical GAN network architecture consists of a generator and discriminator. The generator and discriminator are based on the CNN network architecture.The algorithm of deep learning for image synthesis with the help ofgenerativeadversarialnet-worksis analyzed in the work. During the experiments, the following problems were solved. To increase the initial number of train-ingdata to the datasetapplied a set of affine transformations: mapping, paralleltransfer, shift, scaling, etc. Each of the architectures was trainedfor a certain numberof iterations. The selected architectures were compared by the training timeand image quality based on FID(FreshetInception Distance)metric. The experiments were implemented in Python language.Pytorch was used as a machine learning framework. Based on the used softwarea prototype software module for the synthesis of cytological imageswas developed. Synthesis of cytological images was performed on the basis of DCGAN, WGAN, WGAN-GP, BGAN, BEGAN architectures. Goog-le's online environment called Collaboratory was used for the experimentsusing Nvidia Tesla K80 graphics processor

show abstract

Section: Gan Architectures Overviewmentioning

confidence: 99%

Comparison of generative adversarial networks architectures forbiomedical images synthesis

Berezsky¹,

Liashchynskyi²

2021

AAIT

View full text Add to dashboard Cite

show abstract

“…By using this strategy, their method can compute the KL and reverse KL divergence simultaneously, which in turn increases the variety of samples. Based on this idea, Zhang et al [31] propose a D2GAN variation with two customized discriminators. Specifically, one discriminator consists of residual blocks to form a deep network aiming to increase the variety of generated samples.…”

Section: A Mode Collapse In Gansmentioning

confidence: 99%

Age-Oriented Face Synthesis With Conditional Discriminator Pool and Adversarial Triplet Loss

Wang

Sánchez

2021

IEEE Trans. on Image Process.

View full text Add to dashboard Cite

The vanilla Generative Adversarial Networks (GAN) are commonly used to generate realistic images depicting aged and rejuvenated faces. However, the performance of such vanilla GANs in the age-oriented face synthesis task is often compromised by the mode collapse issue, which may result in the generation of faces with minimal variations and a poor synthesis accuracy. In addition, recent age-oriented face synthesis methods use the L1 or L2 constraint to preserve the identity information on synthesized faces, which implicitly limits the identity permanence capabilities when these constraints are associated with a trivial weighting factor. In this paper, we propose a method for the age-oriented face synthesis task that achieves a high synthesis accuracy with strong identity permanence capabilities. Specifically, to achieve a high synthesis accuracy, our method tackles the mode collapse issue with a novel Conditional Discriminator Pool (CDP), which consists of multiple discriminators, each targeting one particular age category. To achieve strong identity permanence capabilities, our method uses a novel Adversarial Triplet loss. This loss, which is based on the Triplet loss, adds a ranking operation to further pull the positive embedding towards the anchor embedding resulting in significantly reduced intra-class variances in the feature space. Through extensive experiments, we show that our proposed method outperforms state-of-the-art methods in terms of synthesis accuracy and identity permanence capabilities, qualitatively and quantitatively.

show abstract

“…otherwise, it will be difficult to converge [47]. While training the discriminator, the Generator needs to be constant because it needs to learn to differentiate between the generated and fake data.…”

Section: Training the Ganmentioning

confidence: 99%

“…Previous research [40], [44], [47] suggests that training a traditional GAN is hard. The most common problems are as follows.…”

Section: Generating Data Using Ganmentioning

confidence: 99%

Polymorphic Adversarial DDoS attack on IDS using GAN

Chauhan

Heydari

2020

2020 International Symposium on Networks, Computers and Communications (ISNCC)

View full text Add to dashboard Cite

IDS are essential components in preventing malicious traffic from penetrating networks. IDS have been rapidly enhancing their detection ability using ML algorithms. As a result, attackers look for new methods to evade the IDS. Polymorphic attacks are favorites among the attackers as they can bypass the IDS. GAN is a method proven in generating various forms of data. It is becoming popular among security researchers as it can produce indistinguishable data from the original data. I proposed a model to generate DDoS attacks using a WGAN. I used several techniques to update the attack feature profile and generate polymorphic data. This data will change the feature profile in every cycle to test if the IDS can detect the new version attack data. Simulation results from the proposed model show that by continuous changing of attack profiles, defensive systems that use incremental learning will still be vulnerable to new attacks.

show abstract

On the convergence and mode collapse of GAN

Cited by 34 publications

References 1 publication

Comparison of generative adversarial networks architectures forbiomedical images synthesis

Comparison of generative adversarial networks architectures forbiomedical images synthesis

Age-Oriented Face Synthesis With Conditional Discriminator Pool and Adversarial Triplet Loss

Polymorphic Adversarial DDoS attack on IDS using GAN

Contact Info

Product

Resources

About