Biblio

The success of deep learning based steganography has shifted focus of researchers from traditional steganography approaches to deep learning based steganography. Various deep steganographic models have been developed for improved security, capacity and invisibility. In this work a multi-data deep learning steganography model has been developed using a well known deep learning model called Generative Adversarial Networks (GAN) more specifically using deep convolutional Generative Adversarial Networks (DCGAN). The model is capable of hiding two different messages, meant for two different receivers, inside a single cover image. The proposed model consists of four networks namely Generator, Steganalyzer Extractor1 and Extractor2 network. The Generator hides two secret messages inside one cover image which are extracted using two different extractors. The Steganalyzer network differentiates between the cover and stego images generated by the generator network. The experiment has been carried out on CelebA dataset. Two commonly used distortion metrics Peak signal-to-Noise ratio (PSNR) and Structural Similarity Index Metric (SSIM) are used for measuring the distortion in the stego image The results of experimentation show that the stego images generated have good imperceptibility and high extraction rates.

Existing anomaly detection models show success in detecting abnormal images with generative adversarial networks on the insufficient annotation of anomalous samples. However, existing models cannot accurately identify the anomaly samples which are close to the normal samples. We assume that the main reason is that these methods ignore the diversity of patterns in normal samples. To alleviate the above issue, this paper proposes a novel anomaly detection framework based on generative adversarial network, called ADe-GAN. More concretely, we construct a self-supervised learning task to fully explore the pattern information and latent representations of input images. In model inferring stage, we design a new abnormality score approach by jointly considering the pattern information and reconstruction errors to improve the performance of anomaly detection. Extensive experiments show that the ADe-GAN outperforms the state-of-the-art methods over several real-world datasets.

Generative adversarial networks (GANs) have been increasingly popular among deep learning methods. With many GANs-based models developed since its emergence, among which are conditional generative adversarial networks, progressive growing of generative adversarial networks, Wasserstein generative adversarial networks and so on. These frameworks are currently widely applied in areas such as remote sensing cybersecurity, medical, and architecture. Especially, they have solved problems of cloud removal, semantic segmentation, image-to-image translation and data argumentation in remote sensing. For example, WGANs and ProGANs can be applied in data argumentation, and cGANs can be applied in semantic argumentation and image-to-image translation. This article provides an overview of structures of multiple GANs-based models and what areas they can be applied in remote sensing.

Fuzzing is a popular technique for finding soft-ware vulnerabilities. Despite their success, the state-of-art fuzzers will inevitably produce a large number of low-quality inputs. In recent years, Machine Learning (ML) based selection strategies have reported promising results. However, the existing ML-based fuzzers are limited by the lack of training data. Because the mutation strategy of fuzzing can not effectively generate useful input, it is prohibitively expensive to collect enough inputs to train models. In this paper, propose a generative adversarial networks based solution to generate a large number of inputs to solve the problem of insufficient data. We implement the proposal in the American Fuzzy Lop (AFL), and the experimental results show that it can find more crashes at the same time compared with the original AFL.

Deep learning has been widely applied for jobs involving face inpainting, however, there are usually some problems, such as incoherent inpainting edges, lack of diversity of generated images and other problems. In order to get more feature information and improve the inpainting effect, we therefore propose a Generative Adversarial Network of Skip Connection (SCGAN), which connects the encoder layers and the decoder layers by skip connection in the generator. The coherence and consistency of the image inpainting edges are improved, and the finer features of the image inpainting are refined, simultaneously using the discriminator's local and global double discriminators model. We also employ WGAN-GP loss to enhance model stability during training, prevent model collapse, and increase the variety of inpainting face images. Finally, experiments on the CelebA dataset and the LFW dataset are performed, and the model's performance is assessed using the PSNR and SSIM indices. Our model's face image inpainting is more realistic and coherent than that of other models, and the model training is more reliable.

Speech enhancement is often applied for speech-based systems due to the proneness of speech signals to additive background noise. While speech processing-based methods are traditionally used for speech enhancement, with advancements in deep learning technologies, many efforts have been made to implement them for speech enhancement. Using deep learning, the networks learn mapping functions from noisy data to clean ones and then learn to reconstruct the clean speech signals. As a consequence, deep learning methods can reduce what is so-called musical noise that is often found in traditional speech enhancement methods. Currently, one popular deep learning architecture for speech enhancement is generative adversarial networks (GAN). However, the cross-entropy loss that is employed in GAN often causes the training to be unstable. So, in many implementations of GAN, the cross-entropy loss is replaced with the least-square loss. In this paper, to improve the training stability of GAN using cross-entropy loss, we propose to use deep regret analytic generative adversarial networks (Dragan) for speech enhancements. It is based on applying a gradient penalty on cross-entropy loss. We also employ relativistic rules to stabilize the training of GAN. Then, we applied it to the least square and Dragan losses. Our experiments suggest that the proposed method improve the quality of speech better than the least-square loss on several objective quality metrics.

In recent years, generative adversarial networks (GAN) have become a research hotspot in the field of deep learning. Researchers apply them to the field of anomaly detection and are committed to effectively and accurately identifying abnormal images in practical applications. In anomaly detection, traditional supervised learning algorithms have limitations in training with a large number of known labeled samples. Therefore, the anomaly detection model of unsupervised learning GAN is the research object for discussion and research. Firstly, the basic principles of GAN are introduced. Secondly, several typical GAN-based anomaly detection models are sorted out in detail. Then by comparing the similarities and differences of each derivative model, discuss and summarize their respective advantages, limitations and application scenarios. Finally, the problems and challenges faced by GAN in anomaly detection are discussed, and future research directions are prospected.

As one of the defensive solutions against cyberattacks, an Intrusion Detection System (IDS) plays an important role in observing the network state and alerting suspicious actions that can break down the system. There are many attempts of adopting Machine Learning (ML) in IDS to achieve high performance in intrusion detection. However, all of them necessitate a large amount of labeled data. In addition, labeling attack data is a time-consuming and expensive human-labor operation, it makes existing ML methods difficult to deploy in a new system or yields lower results due to a lack of labels on pre-trained data. To address these issues, we propose a semi-supervised IDS model that leverages Generative Adversarial Networks (GANs) and Adversarial AutoEncoder (AAE), called a semi-supervised adversarial autoencoder (SAAE). Our SAAE experimental results on two public datasets for benchmarking ML-based IDS, including NF-CSE-CIC-IDS2018 and NF-UNSW-NB15, demonstrate the effectiveness of AAE and GAN in case of using only a small number of labeled data. In particular, our approach outperforms other ML methods with the highest detection rates in spite of the scarcity of labeled data for model training, even with only 1% labeled data.

Imbalanced class distribution can cause information loss and missed/false alarms for deep learning and machine-learning algorithms. The detection performance of traditional intrusion detection systems tend to degenerate due to skewed class distribution caused by the uneven allocation of observations in different kinds of attacks. To combat class imbalance and improve network intrusion detection performance, we adopt the conditional generative adversarial network (CTGAN) that enables the generation of samples of specific classes of interest. CTGAN builds on the generative adversarial networks (GAN) architecture to model tabular data and generate high quality synthetic data by conditionally sampling rows from the generated model. Oversampling using CTGAN adds instances to the minority class such that both data in the majority and the minority class are of equal distribution. The generated security alerts are used for training classifiers that realize critical alert detection. The proposed scheme is evaluated on a real-world dataset collected from security operation center of a large enterprise. The experiment results show that detection accuracy can be substantially improved when CTGAN is adopted to produce a balanced security-alert dataset. We believe the proposed CTGAN-based approach can cast new light on building effective systems for critical alert detection with reduced missed/false alarms.

With the progress of cryptography computer science, designing cryptographic algorithms using deep learning is a very innovative research direction. Google Brain designed a communication model using generation adversarial network and explored the encrypted communication algorithm based on machine learning. However, the encrypted communication model it designed lacks quantitative evaluation. When some plaintexts and keys are leaked at the same time, the security of communication cannot be guaranteed. This model is optimized to enhance the security by adjusting the optimizer, modifying the activation function, and increasing batch normalization to improve communication speed of optimization. Experiments were performed on 16 bits and 64 bits plaintexts communication. With plaintext and key leak rate of 0.75, the decryption error rate of the decryptor is 0.01 and the attacker can't guess any valid information about the communication.

Learning a disentangled representation of the latent space has become one of the most fundamental problems studied in computer vision. Recently, many Generative Adversarial Networks (GANs) have shown promising results in generating high fidelity images. However, studies to understand the semantic layout of the latent space of pre-trained models are still limited. Several works train conditional GANs to generate faces with required semantic attributes. Unfortunately, in these attempts, the generated output is often not as photo-realistic as the unconditional state-of-the-art models. Besides, they also require large computational resources and specific datasets to generate high fidelity images. In our work, we have formulated a Markov Decision Process (MDP) over the latent space of a pre-trained GAN model to learn a conditional policy for semantic manipulation along specific attributes under defined identity bounds. Further, we have defined a semantic age manipulation scheme using a locally linear approximation over the latent space. Results show that our learned policy samples high fidelity images with required age alterations, while preserving the identity of the person.

As Internet technology gradually matures, the network structure becomes more complex. Therefore, the attack methods of malicious attackers are more diverse and change faster. Fortunately, due to the substantial increase in computer computing power, machine learning is valued and widely used in various fields. It has also been applied to intrusion detection systems. This study found that due to the imperfect data ratio of the unbalanced flow dataset, the model will be overfitting and the misjudgment rate will increase. In response to this problem, this research proposes to use the Cuckoo system to induce malicious samples to generate malicious traffic, to solve the data proportion defect of the unbalanced traffic dataset.

3D reconstruction has piqued the interest of many disciplines, and many researchers have spent the last decade striving to improve on latest automated three-dimensional reconstruction systems. Three Dimensional models can be utilized to tackle a wide range of visualization problems as well as other activities. In this paper, we have implemented a method of Digital Surface Map (DSM) generation from Aerial images using Conditional Generative Adversarial Networks (c-GAN). We have used Seg-net architecture of Convolutional Neural Network (CNN) to segment the aerial images and then the U-net generator of c-GAN generates final DSM. The dataset we used is ISPRS Potsdam-Vaihingen dataset. We also review different stages if 3D reconstruction and how Deep learning is now being widely used to enhance the process of 3D data generation. We provide binary cross entropy loss function graph to demonstrate stability of GAN and CNN. The purpose of our approach is to solve problem of DSM generation using Deep learning techniques. We put forth our method against other latest methods of DSM generation such as Semi-global Matching (SGM) and infer the pros and cons of our approach. Finally, we suggest improvements in our methods that might be useful in increasing the accuracy.

With the advancement in Generative Adversarial Networks (GAN), it has become easier than ever to generate fake images. These images are more realistic and non-discernible by untrained eyes and can be used to propagate fake information on the Internet. In this paper, we propose a novel method to detect GAN generated fake images by using a combination of frequency spectrum of image and deep learning. We apply Discrete Fourier Transform to each of 3 color channels of the image to obtain its frequency spectrum which shows if the image has been upsampled, a common trend in most GANs, and then train a Capsule Network model with it. Conducting experiments on a dataset of almost 1000 images based on Unconditional data modeling (StyleGan2 - ADA) gave results indicating that the model is promising with accuracy over 99% when trained on the state-of-the-art GAN model. In theory, our model should give decent results when trained with one dataset and tested on another.

We are interested in the design of generative networks. The training of these mathematical structures is mostly performed with the help of adversarial (min-max) optimization problems. We propose a simple methodology for constructing such problems assuring, at the same time, consistency of the corresponding solution. We give characteristic examples developed by our method, some of which can be recognized from other applications, and some are introduced here for the first time. We present a new metric, the likelihood ratio, that can be employed online to examine the convergence and stability during the training of different Generative Adversarial Networks (GANs). Finally, we compare various possibilities by applying them to well-known datasets using neural networks of different configurations and sizes.

Document-level relation extraction (RE) aims to extract relations among entities within a document, which is more complex than its sentence-level counterpart, especially in biomedical text mining. Chemical-disease relation (CDR) extraction aims to extract complex semantic relationships between chemicals and diseases entities in documents. In order to identify the relations within and across multiple sentences at the same time, existing methods try to build different document-level heterogeneous graph. However, the entity relation representations captured by these models do not make full use of the document information and disregard the noise introduced in the process of integrating various information. In this paper, we propose a novel model DAM-GAN to document-level biomedical RE, which can extract entity-level and mention-level representations of relation instances with R-GCN and Dual-Attention Multi-Instance Learning (DAM) respectively, and eliminate the noise with Generative Adversarial Network (GAN). Entity-level representations of relation instances model the semantic information of all entity pairs from the perspective of the whole document, while the mention-level representations from the perspective of mention pairs related to these entity pairs in different sentences. Therefore, entity- and mention-level representations can be better integrated to represent relation instances. Experimental results demonstrate that our model achieves superior performance on public document-level biomedical RE dataset BioCreative V Chemical Disease Relation(CDR).

How to obtain accurate travel time predictions is among the most critical problems in Intelligent Transportation Systems (ITS). Recent literature has shown the effectiveness of machine learning models on travel time forecasting problems. However, most of these models predict travel time in a point estimation manner, which is not suitable for real scenarios. Instead of a determined value, the travel time within a future time period is a distribution. Besides, they all use grid structure data to obtain the spatial dependency, which does not reflect the traffic network's actual topology. Hence, we propose GCGTTE to estimate the travel time in a distribution form with Graph Deep Learning and Generative Adversarial Network (GAN). We convert the data into a graph structure and use a Graph Neural Network (GNN) to build its spatial dependency. Furthermore, GCGTTE adopts GAN to approximate the real travel time distribution. We test the effectiveness of GCGTTE with other models on a real-world dataset. Thanks to the fine-grained spatial dependency modeling, GCGTTE outperforms the models that build models on a grid structure data significantly. Besides, we also compared the distribution approximation performance with DeepGTT, a Variational Inference-based model which had the state-of-the-art performance on travel time estimation. The result shows that GCGTTE outperforms DeepGTT on metrics and the distribution generated by GCGTTE is much closer to the original distribution.

Using Generative Adversarial Networks (GAN) to translate images is a significant field in computer vision. There are partial distortion, artifacts and detail loss in the images generated by current image translation algorithms. In order to solve this problem, this paper adds attention-based residual neural network to the generator of GAN. Attention-based residual neural network can improve the representation ability of the generator by weighting the channels of the feature map. Experiment results on the Facades dataset show that Attention Residual GAN can translate images with excellent quality.

Currently, the increasing complexity of DDoS attacks makes it difficult for modern security systems to track them. Machine learning techniques are increasingly being used in such systems as they are well established. However, a new problem arose: the creation of informative datasets. Generative adversarial networks can help create large, high-quality datasets for machine learning training. The article discusses the issue of using generative adversarial networks to generate new patterns of network attacks for the purpose of their further use in training.

Image style transfer refers to the transformation of the style of image, so that the image details are retained to the maximum extent while the style is transferred. Aiming at the problem of low clarity of style transfer images generated by CycleGAN network, this paper improves the CycleGAN network. In this paper, the network model of auto-encoder and variational auto-encoder is added to the structure. The encoding part of the auto-encoder is used to extract image content features, and the variational auto-encoder is used to extract style features. At the same time, the generating network of the model in this paper uses first to adjust the image size and then perform the convolution operation to replace the traditional deconvolution operation. The discriminating network uses a multi-scale discriminator to force the samples generated by the generating network to be more realistic and approximate the target image, so as to improve the effect of image style transfer.

Nowadays, with the increasing use of computers and the Internet, more people are exposed to cyber-security dangers. According to antivirus companies, malware is one of the most common threats of using the Internet. Therefore, providing a practical solution is critical. Current methods use machine learning approaches to classify malware samples automatically. Despite the success of these approaches, the accuracy and efficiency of these techniques are still inadequate, especially for multiple class classification problems and imbalanced training data sets. To mitigate this problem, we use deep learning-based algorithms for classification and generation of new malware samples. Our model is based on the opcode sequences, which are given to the model without any pre-processing. Besides, we use a novel generative adversarial network to generate new opcode sequences for oversampling minority classes. Also, we propose the model that is a combination of Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) to classify malware samples. CNN is used to consider short-term dependency between features; while, LSTM is used to consider longer-term dependence. The experiment results show our method could classify malware to their corresponding family effectively. Our model achieves 98.99% validation accuracy.

While machine learning technologies have been remarkably advanced over the past several years, one of the fundamental requirements for the success of learning-based approaches would be the availability of high-quality data that thoroughly represent individual classes in a problem space. Unfortunately, it is not uncommon to observe a significant degree of class imbalance with only a few instances for minority classes in many datasets, including network traffic traces highly skewed toward a large number of normal connections while very small in quantity for attack instances. A well-known approach to addressing the class imbalance problem is data augmentation that generates synthetic instances belonging to minority classes. However, traditional statistical techniques may be limited since the extended data through statistical sampling should have the same density as original data instances with a minor degree of variation. This paper takes a learning-based approach to data augmentation to enable effective network anomaly detection. One of the critical challenges for the learning-based approach is the mode collapse problem resulting in a limited diversity of samples, which was also observed from our preliminary experimental result. To this end, we present a novel "Divide-Augment-Combine" (DAC) strategy, which groups the instances based on their characteristics and augments data on a group basis to represent a subset independently using a generative adversarial model. Our experimental results conducted with two recently collected public network datasets (UNSW-NB15 and IDS-2017) show that the proposed technique enhances performances up to 21.5% for identifying network anomalies.

Machine learning techniques help to understand underlying patterns in datasets to develop defense mechanisms against cyber attacks. Multilayer Perceptron (MLP) technique is a machine learning technique used in detecting attack vs. benign data. However, it is difficult to construct any effective model when there are imbalances in the dataset that prevent proper classification of attack samples in data. In this research, we use UGR'16 dataset to conduct data wrangling initially. This technique helps to prepare a test set from the original dataset to train the neural network model effectively. We experimented with a series of inputs of varying sizes (i.e. 10000, 50000, 1 million) to observe the performance of the MLP neural network model with distribution of features over accuracy. Later, we use Generative Adversarial Network (GAN) model that produces samples of different attack labels (e.g. blacklist, anomaly spam, ssh scan) for balancing the dataset. These samples are generated based on data from the UGR'16 dataset. Further experiments with MLP neural network model shows that a balanced attack sample dataset, made possible with GAN, produces more accurate results than an imbalanced one.

With the increasing scale of network, attacks against network emerge one after another, and security problems become increasingly prominent. Network intrusion detection system is a widely used and effective security means at present. In addition, with the development of machine learning technology, various intelligent intrusion detection algorithms also start to sprout. By flexibly combining these intelligent methods with intrusion detection technology, the comprehensive performance of intrusion detection can be improved, but the vulnerability of machine learning model in the adversarial environment can not be ignored. In this paper, we study the defense problem of network intrusion detection system against adversarial samples. More specifically, we design a defense algorithm for NIDS against adversarial samples by using bidirectional generative adversarial network. The generator learns the data distribution of normal samples during training, which is an implicit model reflecting the normal data distribution. After training, the adversarial sample detection module calculates the reconstruction error and the discriminator matching error of sample. Then, the adversarial samples are removed, which improves the robustness and accuracy of NIDS in the adversarial environment.

Generative Adversarial Network (GAN) and its variants have shown promising results in generating synthetic data. However, the issues with GANs are: (i) the learning happens around the training samples and the model often ends up remembering them, consequently, compromising the privacy of individual samples - this becomes a major concern when GANs are applied to training data including personally identifiable information, (ii) the randomness in generated data - there is no control over the specificity of generated samples. To address these issues, we propose imdpGAN-an information maximizing differentially private Generative Adversarial Network. It is an end-to-end framework that simultaneously achieves privacy protection and learns latent representations. With experiments on MNIST dataset, we show that imdpGAN preserves the privacy of the individual data point, and learns latent codes to control the specificity of the generated samples. We perform binary classification on digit pairs to show the utility versus privacy trade-off. The classification accuracy decreases as we increase privacy levels in the framework. We also experimentally show that the training process of imdpGAN is stable but experience a 10-fold time increase as compared with other GAN frameworks. Finally, we extend imdpGAN framework to CelebA dataset to show how the privacy and learned representations can be used to control the specificity of the output.