• 정보처리학회논문지B
• /
• 제9B권6호
• /
• pp.735-744
• /
• 2002

본 논문에서는 커널완화법과 동적모멘트를 이용한 support vector machines의 학습성능 개선을 제안하였다. 제안된 학습 방법은 기존의 정적모멘트와는 달리 수렴 정도에 따라 현재의 학습에 과거의 학습 속성을 반영하는 동적모멘트 방법이다. 기존의 정적 상수로 정의된 모멘트가 전체 학습에 동등하게 영향을 주는 반면 제안된 동적모멘트를 이용한 학습방법은 학습 수행에 따라 동적으로 모멘트를 변경함으로써 수렴속도와 학습 성능을 효과적으로 제어할 수 있다. 제안된 학습법을 support vector machine의 새로운 순차 학습 방법인 커널완화법에 적용하였다. 신경망 분류기 표준 평가 데이터인 SONAR 데이터를 이용하여 실험한 곁과 동적모멘트를 이용한 방법이 수렴속도와 학습 성능면에서 기존의 커널완화법과 정적모멘트를 이용한 학습법에 비해 향상된 성능을 보이는 것을 확인하였다.

• 정보처리학회논문지:소프트웨어 및 데이터공학
• /
• 제8권10호
• /
• pp.411-420
• /
• 2019

멀티미디어 포렌식 분야에서 영상을 촬영한 카메라 모델 판별을 위한 연구가 지속되어 왔다. 점점 고도화되는 범죄 중에 불법 촬영 등의 범죄는 카메라가 소형화됨에 따라 피해자가 알아차리기 어렵기 때문에 높은 범죄 발생 건수를 차지하고 있다. 따라서 특정 영상이 어느 카메라로 촬영되었는지를 특정할 수 있는 기술이 사용된다면 범죄자가 자신의 범죄 행위를 부정할 때, 범죄 혐의를 입증할 증거로 사용될 수 있을 것이다. 본 논문에서는 영상을 촬영한 카메라 모델 판별을 위한 딥러닝 모델을 제안한다. 제안하는 모델은 4개의 컨볼루션 계층과 2개의 전연결 계층으로 구성되었으며, 데이터 전처리를 위한 필터로 High Pass Filter를 사용하였다. 제안한 모델의 성능 검증을 위하여 Dresden Image Database를 활용하였고, 데이터셋은 순차분할 방식을 적용하여 생성하였다. 제안하는 모델을 3 계층 모델과 GLCM 적용 모델 등 기존 연구들과 비교 분석을 수행하여 우수성을 보였고, 최신 연구 결과에서 제시하는 수준의 98% 정확도를 달성하였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제17권3호
• /
• pp.958-979
• /
• 2023

Due to laws, regulations, privacy, etc., between 70-90 percent of providers do not share medical data, forming a "data island". It is essential to collaborate across multiple institutions without sharing patient data. Most existing methods adopt distributed learning and centralized federal architecture to solve this problem, but there are problems of resource heterogeneity and data heterogeneity in the practical application process. This paper proposes a collaborative deep learning modelling method based on the blockchain network. The training process uses encryption parameters to replace the original remote source data transmission to protect privacy. Hyperledger Fabric blockchain is adopted to realize that the parties are not restricted by the third-party authoritative verification end. To a certain extent, the distrust and single point of failure caused by the centralized system are avoided. The aggregation algorithm uses the FedProx algorithm to solve the problem of device heterogeneity and data heterogeneity. The experiments show that the maximum improvement of segmentation accuracy in the collaborative training mode proposed in this paper is 11.179% compared to local training. In the sequential training mode, the average accuracy improvement is greater than 7%. In the parallel training mode, the average accuracy improvement is greater than 8%. The experimental results show that the model proposed in this paper can solve the current problem of centralized modelling of multicenter data. In particular, it provides ideas to solve privacy protection and break "data silos", and protects all data.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권4호
• /
• pp.1887-1898
• /
• 2018

In this paper, a method to classify insider threat activity is introduced. The internal threats help detecting anomalous activity in the procedure performed by the user in an organization. When an anomalous value deviating from the overall behavior is displayed, we consider it as an inside threat for classification as an inside intimidator. To solve the situation, Markov Chain Model is employed. The Markov Chain Model shows the next state value through an arbitrary variable affected by the previous event. Similarly, the current activity can also be predicted based on the previous activity for the insider threat activity. A method was studied where the change items for such state are defined by a transition probability, and classified as detection of anomaly of the inside threat through values for a probability variable. We use the properties of the Markov chains to list the behavior of the user over time and to classify which state they belong to. Sequential data sets were generated according to the influence of n occurrences of Markov attribute and classified by machine learning algorithm. In the experiment, only 15% of the Cert: insider threat dataset was applied, and the result was 97% accuracy except for NaiveBayes. As a result of our research, it was confirmed that the Markov Chain Model can classify insider threats and can be fully utilized for user behavior classification.

• 대한산업공학회지
• /
• 제41권6호
• /
• pp.540-550
• /
• 2015

There have been many attempts to find knowledge from data using conventional statistics, data mining, artificial intelligence, machine learning and pattern recognition. In those research areas, knowledge is approached in two ways. Firstly, researchers discover knowledge represented in general features for universal recognition, and secondly, they discover exceptional and distinctive features. In process mining, an instance is sequential information bounded by case ID, known as process instance. Here, an exceptional process instance can cause a problem in the analysis and discovery algorithm. Hence, in this paper we develop a method to detect the knowledge of exceptional and distinctive features when performing process mining. We propose a method for anomaly detection named Distance-based Anomaly Process Instance Detection (DAPID) which utilizes distance between process instances. DAPID contributes to a discovery of distinctive characteristic of process instance. For verifying the suggested methodology, we discovered characteristics of exceptional situations from log data. Additionally, we experiment on real data from a domestic port terminal to demonstrate our proposed methodology.

• International Journal of Contents
• /
• 제15권4호
• /
• pp.8-15
• /
• 2019

This paper presents a solution of the 'Quick, Draw! Doodle Recognition Challenge' hosted by Google. Doodles are drawings comprised of concrete representational meaning or abstract lines creatively expressed by individuals. In this challenge, a doodle is presented as a sequence of sketches. From the view of at the sketch level, to learn the pattern of strokes representing a doodle, we propose a sequential model stacked with multiple convolution layers and Long Short-Term Memory (LSTM) cells following the attention mechanism [15]. From the view at the image level, we use multiple models pre-trained on ImageNet to recognize the doodle. Finally, an ensemble and a post-processing method using the minimum cost flow algorithm are introduced to combine multiple models in achieving better results. In this challenge, our solutions garnered 11th place among 1,316 teams. Our performance was 0.95037 MAP@3, only 0.4% lower than the winner. It demonstrates that our method is very competitive. The source code for this competition is published at: https://github.com/ngxbac/Kaggle-QuickDraw.

• 정보보호학회논문지
• /
• 제21권6호
• /
• pp.161-169
• /
• 2011

최근 지능형 비디오 감시 시스템은 영상 분석 및 인식기술 등의 보다 진화된 기술 개발을 요구하고 있다. 특히, 비디오 영상에서 객체를 식별하기 위하여 Support Vector Machine(SVM)과 같은 기계학습 알고리즘이 이용된다. 그러나 SVM은 대용량의 데이터를 학습시키기 위하여 많은 계산량이 필요하기 때문에 수행시간을 효율적으로 감소시키기 위하여 병렬처리 기법을 적용할 필요가 있다. 본 논문에서는, 최근 사용이 증가하고 있는 멀티코어 프로세서를 활용한 SVM 학습의 병렬처리 방법을 제안한다. 4-코어 프로세서를 이용한 실험 결과, 제안 방법은 SVM 학습의 순차처리 방법과 비교하여 2.5배 정도 수행시간이 감소됨을 확인하였다.

• Journal of Internet Technology
• /
• 제20권3호
• /
• pp.827-838
• /
• 2019

The recent evolution of mobile devices and services have resulted in such plethora of mobile applications (apps) that users have difficulty finding the ones they wish to use in a given moment. We design an app recommendation system which predicts the app to be executed with high accuracy so that users are able to access their next app conveniently and quickly. We introduce the App-Usage Tracking Feature (ATF), a simple but powerful feature for predicting next app launches, which characterizes each app use from the sequence of previously used apps. In addition, our method can be implemented without compromising the user privacy since it is solely trained on the target user's mobile usage data and it can be conveniently implemented in the individual mobile device because of its less computation-intensive behavior. We provide a comprehensive empirical analysis of the performance and characteristics of our proposed method on real-world mobile usage data. We also demonstrate that our system can accurately predict the next app launches and outperforms the baseline methods such as the most frequently used apps (MFU) and the most recently used apps (MRU).

• Structural Engineering and Mechanics
• /
• 제92권2호
• /
• pp.121-131
• /
• 2024

This paper proposes a fault detection method for blade pitch systems of floating wind turbines using transformer-based deep-learning models. Transformers leverage self-attention mechanisms, efficiently process time-series data, and capture long-term dependencies more effectively than traditional recurrent neural networks (RNNs). The model was trained using normal operational data to detect anomalies through high reconstruction losses when encountering abnormal data. In this study, various fault conditions in a blade pitch system, including environmental load cases, were simulated using a detailed model of a spar-type floating wind turbine, the data collected from these simulations were used to train and test the transformer models. The model demonstrated superior fault-detection capabilities with high accuracy, precision, recall, and F1 scores. The results show that the proposed method successfully identifies faults and achieves high-performance metrics, outperforming existing traditional multi-layer perceptron (MLP) models and long short-term memory-autoencoder (LSTM-AE) models. This study highlights the potential of transformer models for real-time fault detection in wind turbines, contributing to more advanced condition-monitoring systems with minimal human intervention.

• 한국정보통신학회:학술대회논문집
• /
• 한국해양정보통신학회 2010년도 춘계학술대회
• /
• pp.205-208
• /
• 2010

본 논문에서는 Zernike 모멘트와 backpropagation신경망을 이용한 온라인 필기체 숫자 인식 방법을 소개한다. 마우스로 통해 입력된 숫자 정보는 전처리를 통해 시간에 순서적이고, 연속적인 좌표 정보로 변환된다. 전처리된 입력 좌표는 Zernike 모멘트(moment)와 각도 특징(angulation feature)을 이용하여 각 숫자가 가지는 고유의 특징을 만들어 낸다. 이러한 특징은 크기, 모양, 틀어진 정도에 상관없이 항상 일정한 성질을 가진다. 제안된 방법으로 추출된 특징은 패턴 구분을 위해 back propagation 신경망의 입력으로 사용된다. 본 논문은 200개의 필기체 숫자 데이터베이스를 이용하여 실험을 한 결과, 제시된 방법은 적은 학습데이터만으로 학습이 가능할 뿐만 아니라 좋은 인식률을 보여준다.