• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2022년도 추계학술발표대회
• /
• pp.724-726
• /
• 2022

딥페이크(deepfake)로 인한 디지털 범죄는 날로 교묘해지면서 사회적으로 큰 파장을 불러일으키고 있다. 이때, 딥러닝 기반 모델의 오류를 발생시키는 적대적 공격(adversarial attack)의 등장으로 딥페이크를 탐지하는 모델의 취약성이 증가하고 있고, 이는 매우 치명적인 결과를 초래한다. 본 연구에서는 2 가지 방법을 통해 적대적 공격에도 영향을 받지 않는 강인한(robust) 모델을 구축하는 것을 목표로 한다. 모델 강화 기법인 적대적 학습(adversarial training)과 영상처리 기반 방어 기법인 크기 변환(resizing), JPEG 압축을 통해 적대적 공격에 대한 강인성을 입증한다.

• 한국BIM학회 논문집
• /
• 제9권1호
• /
• pp.42-51
• /
• 2019

Maintenance of aging structures has attracted societal attention. Maintenance of the aging structure can be efficiently performed with a digital twin. In order to maintain the structure based on the digital twin, it is required to accurately detect the damage of the structure. Meanwhile, deep learning-based damage detection approaches have shown good performance for detecting damage of structures. However, in order to develop such deep learning-based damage detection approaches, it is necessary to use a large number of data before and after damage, but there is a problem that the amount of data before and after the damage is unbalanced in reality. In order to solve this problem, this study proposed a method based on Generative adversarial network, one of Generative Model, for generating acceleration data usually used for damage detection approaches. As results, it is confirmed that the acceleration data generated by the GAN has a very similar pattern to the acceleration generated by the simulation with structural analysis software. These results show that not only the pattern of the macroscopic data but also the frequency domain of the acceleration data can be reproduced. Therefore, these findings show that the GAN model can analyze complex acceleration data on its own, and it is thought that this data can help training of the deep learning-based damage detection approaches.

• 방송공학회논문지
• /
• 제27권3호
• /
• pp.437-450
• /
• 2022

생성적 적대 신경망(GAN, Generative Adversarial Networks)는 이미지 생성 분야에서 주목할 만한 발전을 이루었다. 하지만 큰 데이터 셋에서 불안정한 모습을 보인다는 한계 때문에 다양한 응용 분야에 쉽게 적용하기 어렵다. 단일 이미지 생성적 적대 신경망은 한장의 이미지의 내부 분포를 잘 학습하여 다양한 영상을 생성하는 분야이다. 큰 데이터셋이 아닌 단 한장만 학습함으로써 안정적인 학습이 가능하며 이미지 리타겟팅, 이미지 조작, super resolution 등 다양한 분야에 활용 가능하다. 본 논문에서는 SinGAN, ConSinGAN, InGAN, DeepSIM, 그리고 One-Shot GAN 총 다섯 개의 단일 이미지 생성적 적대 신경망을 살펴본다. 우리는 각각의 단일 이미지 생성적 적대 신경망 모델들의 성능을 비교하고 장단점을 분석한다.

• International Journal of Computer Science & Network Security
• /
• 제22권9호
• /
• pp.23-30
• /
• 2022

Glaucoma is a chronic neuropathy that affects the optic nerve which can lead to blindness. The detection and prediction of glaucoma become possible using deep neural networks. However, the detection performance relies on the availability of a large number of data. Therefore, we propose different frameworks, including a hybrid of a generative adversarial network and a convolutional neural network to automate and increase the performance of glaucoma detection. The proposed frameworks are evaluated using five public glaucoma datasets. The framework which uses a Deconvolutional Generative Adversarial Network (DCGAN) and a DenseNet pre-trained model achieves 99.6%, 99.08%, 99.4%, 98.69%, and 92.95% of classification accuracy on RIMONE, Drishti-GS, ACRIMA, ORIGA-light, and HRF datasets respectively. Based on the experimental results and evaluation, the proposed framework closely competes with the state-of-the-art methods using the five public glaucoma datasets without requiring any manually preprocessing step.

• 정보처리학회논문지:소프트웨어 및 데이터공학
• /
• 제8권6호
• /
• pp.257-264
• /
• 2019

최근 자연언어처리 분야에서 딥러닝 모델이 좋은 성과를 보이고 있다. 이러한 딥러닝 모델의 성능을 향상시키기 위해서는 많은 양의 데이터가 필요하다. 하지만 많은 양의 데이터를 모으기 위해서는 많은 인력과 시간이 소요되기 때문에 데이터 확장을 통해 이와 같은 문제를 해소할 수 있다. 그러나 문장 데이터의 경우 이미지 데이터에 비해 데이터 변형이 어렵기 때문에 다양한 문장을 생성할 수 있는 생성 모델을 통해 문장 데이터 자동 확장을 해보고자 한다. 본 연구에서는 최근 이미지 생성 모델에서 좋은 성능을 보이고 있는 생성적 적대 신경망 중 하나인 CS-GAN을 사용하여 학습 데이터로부터 새로운 문장들을 생성해 보고 유용성을 다양한 지표로 평가하였다. 평가 결과 CS-GAN이 기존의 언어 모델을 사용할 때보다 다양한 문장을 생성할 수 있었고 생성된 문장을 감성 분류기에 학습시켰을 때 감성 분류기의 성능이 향상됨을 보였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권11호
• /
• pp.5594-5615
• /
• 2019

Malicious social robots, which are disseminators of malicious information on social networks, seriously affect information security and network environments. The detection of malicious social robots is a hot topic and a significant concern for researchers. A method based on classification has been widely used for social robot detection. However, this method of classification is limited by an unbalanced data set in which legitimate, negative samples outnumber malicious robots (positive samples), which leads to unsatisfactory detection results. This paper proposes the use of generative adversarial networks (GANs) to extend the unbalanced data sets before training classifiers to improve the detection of social robots. Five popular oversampling algorithms were compared in the experiments, and the effects of imbalance degree and the expansion ratio of the original data on oversampling were studied. The experimental results showed that the proposed method achieved better detection performance compared with other algorithms in terms of the F1 measure. The GAN method also performed well when the imbalance degree was smaller than 15%.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권11호
• /
• pp.5427-5445
• /
• 2019

While deep neural networks have achieved remarkable performance in representation learning, a huge amount of labeled training data are usually required by supervised deep models such as convolutional neural networks. In this paper, we propose a new representation learning method, namely generative adversarial networks (GAN) based bagging deep convolutional autoencoders (GAN-BDCAE), which can map data to diverse hierarchical representations in an unsupervised fashion. To boost the size of training data, to train deep model and to aggregate diverse learning machines are the three principal avenues towards increasing the capabilities of representation learning of neural networks. We focus on combining those three techniques. To this aim, we adopt GAN for realistic unlabeled sample generation and bagging deep convolutional autoencoders (BDCAE) for robust feature learning. The proposed method improves the discriminative ability of learned feature embedding for solving subsequent pattern recognition problems. We evaluate our approach on three standard benchmarks and demonstrate the superiority of the proposed method compared to traditional unsupervised learning methods.

• 전자통신동향분석
• /
• 제35권4호
• /
• pp.91-102
• /
• 2020

Recently, generative adversarial networks (GANs) is a field of research that has rapidly emerged wherein many studies conducted shows overwhelming results. Initially, this was at the level of imitating the training dataset. However, the GAN is currently useful in many fields, such as transformation of data categories, restoration of erased parts of images, copying facial expressions of humans, and creation of artworks depicting a dead painter's style. Although many outstanding research achievements have been attracting attention recently, GANs have encountered many challenges. First, they require a large memory facility for research. Second, there are still technical limitations in processing high-resolution images over 4K. Third, many GAN learning methods have a problem of instability in the training stage. However, recent research results show images that are difficult to distinguish whether they are real or fake, even with the naked eye, and the resolution of 4K and above is being developed. With the increase in image quality and resolution, many applications in the field of design and image and video editing are now available, including those that draw a photorealistic image as a simple sketch or easily modify unnecessary parts of an image or a video. In this paper, we discuss how GANs started, including the base architecture and latest technologies of GANs used in high-resolution, high-quality image creation, image and video editing, style translation, content transfer, and technology.

• International Journal of Internet, Broadcasting and Communication
• /
• 제11권4호
• /
• pp.37-42
• /
• 2019

In this paper, we explore the details of three classic data augmentation methods and two generative model based oversampling methods. The three classic data augmentation methods are random sampling (RANDOM), Synthetic Minority Over-sampling Technique (SMOTE), and Adaptive Synthetic Sampling (ADASYN). The two generative model based oversampling methods are Conditional Generative Adversarial Network (CGAN) and Wasserstein Generative Adversarial Network (WGAN). In imbalanced data, the whole instances are divided into majority class and minority class, where majority class occupies most of the instances in the training set and minority class only includes a few instances. Generative models have their own advantages when they are used to generate more plausible samples referring to the distribution of the minority class. We also adopt CGAN to compare the data augmentation performance with other methods. The experimental results show that WGAN-based oversampling technique is more stable than other approaches (RANDOM, SMOTE, ADASYN and CGAN) even with the very limited training datasets. However, when the imbalanced ratio is too small, generative model based approaches cannot achieve satisfying performance than the conventional data augmentation techniques. These results suggest us one of future research directions.

• Journal of information and communication convergence engineering
• /
• 제19권4호
• /
• pp.228-233
• /
• 2021

In this study, we analyze the credit information (loan, delinquency information, etc.) of individual business owners to generate voluminous training data to establish a bankruptcy prediction model through a partial synthetic training technique. Furthermore, we evaluate the prediction performance of the newly generated data compared to the actual data. When using conditional tabular generative adversarial networks (CTGAN)-based training data generated by the experimental results (a logistic regression task), the recall is improved by 1.75 times compared to that obtained using the actual data. The probability that both the actual and generated data are sampled over an identical distribution is verified to be much higher than 80%. Providing artificial intelligence training data through data synthesis in the fields of credit rating and default risk prediction of individual businesses, which have not been relatively active in research, promotes further in-depth research efforts focused on utilizing such methods.