China
Africa
Explore chapters and articles related to this topic
Privacy Preservation with Machine Learning
Published in Sudhir Kumar Sharma, Bharat Bhushan, Narayan C. Debnath, IoT Security Paradigms and Applications, 2020
P. Bhuvaneswari, Nagender Kumar Suryadevara
In [37], the authors used the WGAN (Wasserstein GAN) model and proved theoretically that their model with stochastic gradient descent (SGD) optimizer helps to improve the quality of generated data when applied on numerical data. To test the effectiveness of WGAN, WGAN is compared with GAN encoder–decoder and DP. The summary of comparisons using experiment result analysis is as follows: (a) All three models can preserve data privacy by masking the entire original data; (b) the generated data is different from the original data in all three compared models; (c) but only WGAN statistical characteristics are the same as the original dataset and the statistical characteristics of the remaining two models are different from the actual dataset. The experiment uses simulated datasets (for the census and the environmental simulation for training of model and testing).
Applications of Physics-Informed Scientific Machine Learning in Subsurface Science: A Survey
Published in Anuj Karpatne, Ramakrishnan Kannan, Vipin Kumar, Knowledge-Guided Machine Learning, 2023
Alexander Y. Sun, Hongkyu Yoon, Chung-Yan Shih, Zhi Zhong
Many variants of the vanilla GAN have been proposed, such as the deep convolutional GAN (DCGAN)(Radford, Metz, and Chintala, 2015), superresolution GAN (SRGAN) (Ledig et al., 2017), Cycle-GAN (Zhu et al., 2017), StarGAN (Choi et al., 2018), and missing data imputation GAN (GAIN) (Yoon, Jordon, and Van Der Schaar, 2018). Recent surveys of GANs are provided in (Creswell et al., 2018; Pan et al., 2019). So far, GANs have demonstrated superb performance in generating photo-realistic images and learning cross-domain mappings. Training of the GANs, however, are known to be challenging due to (a) larger-size networks, especially those involving a long chain of CNN blocks and multiple pairs of generators/discriminators, (b) the nonconvex cost functions used in GAN formulations, (c) diminished gradient issue, namely, the discriminator is trained so well early on in training that the generator's gradient vanishes and learns nothing, and (d) the “mode collapse” problem, namely, the generator only returns samples from a small number of modes of a mutimodal distribution (Goodfellow et al., 2016). In the literature, different strategies have been proposed to alleviate some of the aforementioned issues. For example, to adopt and modify deep CNNs for improving training stability, the DCGAN architecture (Radford, Metz, and Chintala, 2015) was proposed by including stride convolutions and ReLu/LeakyRelu activation functions in the convolution layers. To ameliorate stability issues with the GAN loss function, the Wasserstein distance was introduced in the Wasserstein GAN (WGAN) (Arjovsky, Chintala, Bottou, 2017; Gulrajani et al., 2017) to measure the distance between generated and real data samples, which was then used as the training criterion in a critic model. To remedy the mode collapse problem, the multimodal GAN (Huang et al., 2018) was introduced, in which the latent space is assumed to consist of domain-invariant (called content code) and domain specific (called style code) parts; the former is shared by all domains, while the latter is only specific to one domain. The multimodal GAN is trained by minimizing the image space reconstruction loss, and the latent space reconstruction loss. In the context of continual learning, the memory replay GAN (Wu et al., 2018) was proposed to learn from a sequence of disjoint tasks. Like the autoencoders, GAN represents a general formulation for supervised and semi-supervised learning, thus its implementation is not restricted to certain types of network models.
A Review on Application of GANs in Cybersecurity Domain
Published in IETE Technical Review, 2022
Steganography is a technique designed to protect confidential and sensitive information from malicious attacks, by embedding it into an audio, video, image, or text file. Similarly, Image Steganography refers to hide the data within an image where the image selected for this purpose is called as cover-image, and the image obtained after steganography is called as stego-image. The authors of [14] reviewed specific strategies for steganography using GANs such as cover selection, modification, and synthesis. Secure Steganography based on GANs (SSGAN) [15] is proposed to generate suitable and secure covers for steganography. The model built upon Wasserstein GAN (WGAN) consists of a generator network and two discriminator networks (D and S). The generator network is used to generate realistic images as well as secure covers for steganography. The network D is used to evaluate the visual quality of generated images, while the network S is used to judge whether the images generated are suitable and secure covers for steganography. VAE-SGAN [16] proposed a combination of generative adversarial networks and variational auto-encoders (VAE) [17], which follows a similar model architecture as SSGAN but with a new encoder network to generate better visually convincing images with less model collapse to achieve better security in spatial image steganography. The encoder network converts raw images to low dimensional latent vectors with certain characteristic features. The generator, also called decoder converts the latent vector to images close to raw images. The discriminator classifies images as raw or generated, while the steganalyser differentiates between stego images and generated cover images.
Network-wide ride-sourcing passenger demand origin-destination matrix prediction with a generative adversarial network
Published in Transportmetrica A: Transport Science, 2022
Changlin Li, Liang Zheng, Ning Jia
As is known, GAN has been one of the pioneering deep learning technologies (Lv, Chen, and Fei-Yue Wang 2018). Goodfellow et al. (2014) firstly proposed the GAN framework, which covers competition and optimisation between the discriminator and the generator until fake data are indistinguishable from real data. Later, its variants have been developed extensively. For example, Mirza and Osindero (2014) proposed CGAN by inputting conditional information to the generator and the discriminator. By embedding CNN into the original GAN, Radford, Metz, and Chintala (2015) put forward a Deep Convolutional GAN (DCGAN) with better generation quality and training stability. Arjovsky, Chintala, and Bottou (2017) introduced a Wasserstein GAN (WGAN) also with the improved training stability. Gulrajani et al. (2017) improved WGAN by introducing gradient penalty, which results in Wasserstein Generative Adversarial Network with Gradient penalty (WGAN-GP) without a strong continuity condition, i.e. Lipschitz continuity for the entire sample space. As for applying GANs in traffic prediction and estimation, Zhang et al. (2019) proposed Trip Information Maximizing Generative Adversarial Network (T-InfoGAN) to estimate trip travel time distribution by modelling the joint distribution of travel times of two successive links with the consideration of network-wide spatiotemporal correlations. Later, Zhang et al. (2021) presented a travel times imputation generative adversarial network (TTI-GAN) for travel times imputation under various data missing rates, also considering network-wide spatiotemporal correlations. More relevant, considering both spatiotemporal and external factors, Saxena and Cao (2019) proposed a novel deep generative adversarial network (D-GAN) model for taxi demand prediction. However, the Kullback–Leibler (KL) divergence used in the discriminator loss function to measure the distance between the predicted and real demand distributions would easily cause various problems, such as converging difficulty, mode collapse, and vanishing gradient (Goodfellow et al. 2014). Yu et al. (2019) adopted CGAN to predict taxi passenger demand with conditional information such as taxi passenger demand, link length, etc. However, to address data dimensionality and data sparsity of OD matrix prediction (Zhang et al. 2021), it is desirable to design a suitable GAN to capture complicated internal spatiotemporal correlations and external dependencies aforementioned.