• Annual Conference of KIPS
• /
• 2024.05a
• /
• pp.666-669
• /
• 2024

연합학습은 원본 데이터를 공유하지 않고 모델을 학습할 수 있는 각광받는 프라이버시를 위한 학습방법론이다. 이를 위해 참여자의 데이터를 수집하는 대신, 데이터를 인공지능 모델 학습의 요소들(가중치, 기울기 등)로 변환한 뒤, 이를 공유한다. 이러한 강점에 더해 기존 연합학습을 개선하는 방법론들이 추가적으로 연구되고 있다. 기존 연합학습은 모델 가중치를 평균내는 것으로 참여자 간에 동일한 모델 구조를 강요하기 때문에, 참여자 별로 자신의 환경에 알맞은 모델 구조를 사용하기 어렵다. 이를 해결하기 위해 지식 증류 기반의 연합학습 방법(Knowledge Distillation-based Federated Learning)으로 서로 다른 모델 구조를 가질 수 있도록(Model Heterogenousity) 하는 방법이 제시되고 있다. 연합학습은 여러 참여자가 연합하기 때문에 일부 악의적인 참여자로 인한 모델 포이즈닝 공격에 취약하다. 수많은 연구들이 기존 가중치를 기반으로한 연합학습에서의 위협을 연구하였지만, 지식 증류 기반의 연합학습에서는 이러한 위협에 대한 조사가 부족하다. 본 연구에서는 최초로 지식 증류 기반의 연합학습에서의 모델 성능 하락 공격에 대한 위협을 실체화하고자 한다. 이를 위해 우리는 GMA(Gaussian-based Model Poisoning Attack)과 SMA(Sign-Flip based Model Poisoning Attack)을 제안한다. 결과적으로 우리가 제안한 공격 방법은 실험에서 최신 학습 기법에 대해 평균적으로 모델 정확도를 83.43%에서 무작위 추론에 가깝게 떨어뜨리는 것으로 공격 성능을 입증하였다. 우리는 지식 증류 기반의 연합학습의 강건성을 평가하기 위해, 새로운 공격 방법을 제안하였고, 이를통해 현재 지식 증류 기반의 연합학습이 악의적인 공격자에 의한 모델 성능 하락 공격에 취약한 것을 보였다. 우리는 방대한 실험을 통해 제안하는 방법의 성능을 입증하고, 결과적으로 강건성을 높이기 위한 많은 방어 연구가 필요함을 시사한다.

• Journal of the Korean Association of Geographic Information Studies
• /
• v.24 no.1
• /
• pp.92-111
• /
• 2021

As fine dust negatively affects disease, industry and economy, the people are sensitive to fine dust. Therefore, if the occurrence of fine dust can be predicted, countermeasures can be prepared in advance, which can be helpful for life and economy. Fine dust is affected by the weather and the degree of concentration of fine dust emission sources. The industrial sector has the largest amount of fine dust emissions, and in industrial complexes, factories emit a lot of fine dust as fine dust emission sources. This study targets regions with old industrial complexes in local cities. The purpose of this study is to explore the factors that cause fine dust and develop a predictive model that can predict the occurrence of fine dust. weather data and fine dust data were used, and variables that influence the generation of fine dust were extracted through multiple regression analysis. Based on the results of multiple regression analysis, a model with high predictive power was extracted by learning with a machine learning regression learner model. The performance of the model was confirmed using test data. As a result, the models with high predictive power were linear regression model, Gaussian process regression model, and support vector machine. The proportion of training data and predictive power were not proportional. In addition, the average value of the difference between the predicted value and the measured value was not large, but when the measured value was high, the predictive power was decreased. The results of this study can be developed as a more systematic and precise fine dust prediction service by combining meteorological data and urban big data through local government data hubs. Lastly, it will be an opportunity to promote the development of smart industrial complexes.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.66 no.10
• /
• pp.1516-1522
• /
• 2017

We propose a novel monitoring system for diagnosing crack faults of the wind turbine using image information. The proposed method classifies a normal state and a abnormal state for the blade parts of the wind turbine. Specifically, the images are input to the proposed system in various states of wind turbine rotation. according to the blade condition. Then, the video of rotating blades on the wind turbine is divided into several image frames. Motion vectors are estimated using the previous and current images using the motion estimation, and the change of the motion vectors is analyzed according to the blade state. Finally, we determine the final blade state using the Support Vector Machine (SVM) classifier. In SVM, features are constructed using the area information of the blades and the motion vector values. The experimental results showed that the proposed method had high classification performance and its $F_1$ score was 0.9790.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.65 no.6
• /
• pp.1070-1079
• /
• 2016

In this paper, the design of recursive radial basis function neural networks based on incremental fuzzy c-means is introduced for processing the big data. Radial basis function neural networks consist of condition, conclusion and inference phase. Gaussian function is generally used as the activation function of the condition phase, but in this study, incremental fuzzy clustering is considered for the activation function of radial basis function neural networks, which could effectively do big data processing. In the conclusion phase, the connection weights of networks are given as the linear function. And then the connection weights are calculated by recursive least square estimation. In the inference phase, a final output is obtained by fuzzy inference method. Machine Learning datasets are employed to demonstrate the superiority of the proposed classifier, and their results are described from the viewpoint of the algorithm complexity and performance index.

• Smart Media Journal
• /
• v.9 no.3
• /
• pp.80-89
• /
• 2020

Out - of - distribution (OOD) samples are frequently encountered when deploying a classification model in plenty of real-world machine learning-based applications. Those samples are normally sampling far away from the training distribution, but many classifiers still assign them high reliability to belong to one of the training categories. In this study, we address the problem of removing OOD examples by estimating marginal density estimation using variational autoencoder (VAE). We also investigate other proper methods, such as temperature scaling, Gaussian discrimination analysis, and label smoothing. We use Chonnam National University (CNU) weeds dataset as the in - distribution dataset and CIFAR-10, CalTeach as the OOD datasets. Quantitative results show that the proposed framework can reject the OOD test samples with a suitable threshold.

• The Journal of Korea Robotics Society
• /
• v.7 no.3
• /
• pp.194-204
• /
• 2012

Measuring task complexity of movement skill is an important factor to evaluate a difficulty of learning and/or imitating a task for autonomous robots. Although many complexity-measures are proposed in research areas such as neuroscience, physics, computer science, and biology, there have been little attention on the robotic tasks. To cope with measuring complexity of robotic task, we propose an information-theoretic measure for task complexity of movement skills. By modeling proprioceptive as well as exteroceptive sensor data as multivariate Gaussian distribution, movements of a task can be modeled as probabilistic model. Additionally, complexity of temporal variations is modeled by sampling in time and modeling as individual random variables. To evaluate our proposed complexity measure, several experiments are performed on the real robotic movement tasks.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.43 no.2 s.308
• /
• pp.33-39
• /
• 2006

This paper presents the modified LVQ method for performance improvement of pattern classification. The proposed method uses the skewness of probability distribution between the input vectors and the reference vectors. During training, the reference vectors are closest to the input vectors using the probabilistic distribution of the input vectors, and they are positioned to approximate the decision surfaces of the theoretical Bayes classifier. In order to verify the effectiveness of the proposed method, we performed experiments on the Gaussian distribution data set, and the Fisher's IRIS data set. The experimental results show that the proposed method considerably improves on the performance of the LVQ1, LVQ2, and GLVQ.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2010.05a
• /
• pp.981-984
• /
• 2010

In this paper, we used Support Vector Machines(SVMs) as the learning method, one of Artificial Neural Network, to segregated from the continuous speech into phonemes, an initial, medial, and final sound, and then, performed continuous speech recognition from it. A Decision boundary of phoneme is determined by algorithm with maximum frequency in a short interval. Speech recognition process is performed by Continuous Hidden Markov Model(CHMM), and we compared it with another phoneme segregated from the eye-measurement. From the simulation results, we confirmed that the method, SVMs, we proposed is more effective in an initial sound than Gaussian Mixture Models(GMMs).

• Journal of the Korean Society for Industrial and Applied Mathematics
• /
• v.25 no.4
• /
• pp.296-311
• /
• 2021

Topological Data Analysis (TDA), a relatively new field of data analysis, has proved very useful in a variety of applications. The main persistence tool from TDA is persistent homology in which data structure is examined at many scales. Representations of persistent homology include persistence barcodes and persistence diagrams, both of which are not straightforward to reconcile with traditional machine learning algorithms as they are sets of intervals or multisets. The problem of faithfully representing barcodes and persistent diagrams has been pursued along two main avenues: kernel methods and vectorizations. One vectorization is the Betti sequence, or Betti curve, derived from the persistence barcode. While the Betti sequence has been used in classification problems in various applications, to our knowledge, the stability of the sequence has never before been discussed. In this paper we show that the Betti sequence is unstable under the 1-Wasserstein metric with regards to small perturbations in the barcode from which it is calculated. In addition, we propose a novel stabilized version of the Betti sequence based on the Gaussian smoothing seen in the Stable Persistence Bag of Words for persistent homology. We then introduce the normalized cumulative Betti sequence and provide numerical examples that support the main statement of the paper.

• Structural Engineering and Mechanics
• /
• v.71 no.6
• /
• pp.739-749
• /
• 2019

This article deals with the application of reliability analysis for determining the safety of simply supported beam under the uniformly distributed load. The uncertainties of the existing methods were taken into account and hence reliability analysis has been adopted. To accomplish this aim, Generalized Regression Neural Network (GRNN), Extreme Learning Machine (ELM) and Gaussian Process Regression (GPR) models are developed. Reliability analysis is the probabilistic style to determine the possibility of failure free operation of a structure. The application of probabilistic mathematics into the quantitative aspects of a structure and improve the qualitative aspects of a structure. In order to construct the GRNN, ELM and GPR models, the dataset contains Modulus of Elasticity (E), Load intensity (w) and performance function (${\delta}$) in which E and w are inputs and ${\delta}$ is the output. The achievement of the developed models was weighed by various statistical parameters; one among the most primitive parameter is Coefficient of Determination ($R^2$) which has 0.998 for training and 0.989 for testing. The GRNN outperforms the other ELM and GPR models. Other different statistical computations have been carried out, which speaks out the errors and prediction performance in order to justify the capability of the developed models.