Adversarial Detection with Gaussian Process Regression-based Detector

( Sangheon Lee ),( Noo-ri Kim ),( Youngwha Cho ),( Jae-young Choi ),( Suntae Kim ),( Jeong-ah Kim ),( Jee-hyong Lee ) 한국인터넷정보학회 2019 KSII Transactions on Internet and Information Syst Vol.13 No.8

Adversarial attack is a technique that causes a malfunction of classification models by adding noise that cannot be distinguished by humans, which poses a threat to a deep learning model. In this paper, we propose an efficient method to detect adversarial images using Gaussian process regression. Existing deep learning-based adversarial detection methods require numerous adversarial images for their training. The proposed method overcomes this problem by performing classification based on the statistical features of adversarial images and clean images that are extracted by Gaussian process regression with a small number of images. This technique can determine whether the input image is an adversarial image by applying Gaussian process regression based on the intermediate output value of the classification model. Experimental results show that the proposed method achieves higher detection performance than the other deep learning-based adversarial detection methods for powerful attacks. In particular, the Gaussian process regression-based detector shows better detection performance than the baseline models for most attacks in the case with fewer adversarial examples.

Energy consumption model with energy use factors of tenants in commercial buildings using Gaussian process regression

Yoon, Young Ran,Moon, Hyeun Jun Elsevier 2018 Energy and buildings Vol.168 No.-

<P><B>Abstract</B></P> <P>Identification of the factors influencing energy consumption in buildings is crucial for energy efficient control in the operation stage. By using a multi-variate approach in energy performance prediction, we can characterize the building energy usage with a few available variables. However, very few studies have applied the variables related to building operation for energy consumption in buildings. Especially, the importance of the use factor, which affects the energy consumption, varies depending on the usage of tenants. However, there is a lack of sufficient research on the energy consumption based on energy use factors of each tenant in a building.</P> <P>Therefore, in this study we propose an energy consumption model using data on the energy use factors, such as occupant schedule, operation, and equipment, especially with a focus on the tenants in buildings. In this study, we analyzed the ranking of variable importance using the Random Forest algorithm and verified the energy consumption results of individual, office, and retail tenants in commercial buildings using a Gaussian process regression model. The main contribution of this study is the identification of the influence of energy use factors on the energy consumption of each tenant, both office and retail, thereby developing an energy model. This study established a method to identify the combination of variables that could estimate the energy consumption. Moreover, it can be seen that the significant variables to consider for developing an energy model differ depending on the tenant use class, i.e., office or retail.</P> <P><B>Highlights</B></P> <P> <UL> <LI> Identified the influence of the energy use factors (e.g., occupant related factors, indoor air conditions such as heating set temperature, equipment types such as heating system, lighting system, and plugged load) for the energy consumption of each tenant in commercial buildings. </LI> <LI> Established a method to identify the combination of variables that could provide a more accurate estimation of energy consumption using Gaussian process regression. </LI> <LI> Identified that the significant variables to consider for developing an energy consumption model differ depending on the tenant usage, i.e., office or retail. </LI> </UL> </P>

Gaussian Process Approximate Dynamic Programming for Energy Management of Parallel Hybrid Electric Vehicles

Jin Woo Bae,Dohee Kim,Jeongsoo Eo,Kwang-Ki K. Kim 제어로봇시스템학회 2020 제어로봇시스템학회 국제학술대회 논문집 Vol.2020 No.10

This paper presents two non-parametric Bayesian techniques–Gaussian Process Dynamic Programming (GPDP) and Gaussian Process Dynamic Programming-Receding Horizon Control (GPDP-RHC)–for optimal energy management of parallel hybrid electric vehicles. Hybrid electric vehicles (HEVs) are powered by engine and electric machine and assigning the required traction power to the two sources. It is known as the supervisory control which can be formulated as an optimal control problem. To solve the supervisory optimal control, we adopt the approximate dynamic programming (ADP) with Gaussian processes (GPs) which are used for value function approximation in optimal control problem. High-fidelity models of real-world vehicle data for battery, engine, and electric machine are used to obtain discrete dynamic programming (DP) solutions for a known driving cycle. To overcome limitations in real-time application of DP, we use non-parametric Bayesian function approximation techniques using GPs. The state-value tables obtained by dynamic programming are approximated by Gaussian process regression. Furthermore, the future value function is predicted by GPDP in one-step lookahead with RHC. For demonstration of optimality and efficiency, the proposed GPDP-RHC solution is compared with both the offline global DP solution and real-driving result.

적응적 역전파를 이용한 신경망 기반의 가우시안 프로세스 회귀 처리 및 해의 안정성 모델링

양정연 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.12

This paper describes a study on a modified neural network model to cover the Gaussian process regression method, which supports for probabilistic confidence interval of functional distribution and noisy signals. Discrimination function is redesigned to handle deviated errors under normal gaussian distribution. This modified error function is combined with typical back propagation process. Thus, the results become the superposition of weighted Gaussian functions and the decision boundaries, as in the case of conventional Gaussian process regression methods. Owing to the discontinuity of discrimination function, the stability of the proposed error function becomes deteriorated. The concept of low pass filtering is applied to improve the unstable problems of weight updating process. 본 논문은 공간 내 확률적 분포를 가진 함수 또는 잡음 신호를 확률 기반의 신뢰 구간으로 표현하기 위해, 기존의 역전파 신경망 기법을 변형하여 가우시안 프로세스의 회귀에 의한 결정 공간의 추종방식을 제안하고자 한다. 신경망 기법의 역전파 과정에 사용되는 오차 함수를 부호에 따라 가변적 가중치를 가지도록 설계하고, 이를 가우시안 함수에 기반한 표준 편차 처리에 적합하도록 변형함에 따라, 가우시안 확률 분포 함수의 조합에 의한 결정 경계의 추종 및 신경망 노드의 합성에 의한 가우시안 프로세스의 회귀 처리 과정을 수행하고자 한다. 수정된 오차 함수의 경우, 불연속성에 기인한 해 탐색 과정의 불안정성이 존재하기에 이를 회피하기 위해 저역 통과 필터의 개념을 이용한 안정성 방법에 대해 논하고자 한다.

가우시안 프로세스 회귀분석을 이용한 영상초점으로부터의 3차원 형상 재구성

타릭,최영규 한국반도체디스플레이기술학회 2012 반도체디스플레이기술학회지 Vol.11 No.3

The accuracy of Shape From Focus (SFF) technique depends on the quality of the focus measurements which are computed through a focus measure operator. In this paper, we introduce a new approach to estimate 3D shape of an object based on Gaussian process regression. First, initial depth is estimated by applying a conventional focus measure on image sequence and maximizing it in the optical direction. In second step, input feature vectors consisting of eginvalues are computed from 3D neighborhood around the initial depth. Finally, by utilizing these features, a latent function is developed through Gaussian process regression to estimate accurate depth. The proposed approach takes advantages of the multivariate statistical features and covariance function. The proposed method is tested by using image sequences of various objects. Experimental results demonstrate the efficacy of the proposed scheme.

Controlled kinetic Monte Carlo simulation of laser improved nano particle deposition process

Song, Ji-Hyeon,Choi, Kweon-Hoon,Dai, Ruonan,Choi, Jung-Oh,Ahn, Sung-Hoon,Wang, Yan Elsevier Sequoia 2018 Powder Technology Vol. No.

<P><B>Abstract</B></P> <P>Thin film coating is important in many applications such as electrodes, sensors, and energy devices. Nano particle deposition is one of the most used additive manufacturing processes for coating. It has advantages of efficiency, cost effectiveness, and ease of controlling film properties. Recent experimental studies showed that laser can enhance the efficiency of the deposition process. However, there is still a lack of fundamental understanding of the laser treatment effect on nano particles, which makes process control difficult and ad hoc. In this research, the effect of laser treatment on morphological change of films in the nano particle deposition system is studied with controlled kinetic Monte Carlo (cKMC) simulation. cKMC is a generalized version of classical kinetic Monte Carlo, which can be used to simulate both controlled and self-assembly processes at atomistic level with larger sizes and longer time scales than molecular dynamics. In this work, a coarse-grained cKMC model is constructed to simulate diffusion, laser irradiation, and deposition processes simultaneously. The simulation model is calibrated with experimental data. Different laser irradiation conditions on alumina particles are studied, which result in different thickness and porosity of the deposited layers. A Gaussian process regression modeling approach is also developed for model validation with the consideration of observation bias and discrepancy. Simulation results are in good agreement with the experimental results.</P> <P><B>Highlights</B></P> <P> <UL> <LI> Alumina nano particle deposition in aerosol flow is simulated with controlled kinetic Monte Carlo. </LI> <LI> Laser irradiation, diffusion, and deposition processes are simulated efficiently. </LI> <LI> A Gaussian process regression approach is proposed for model validation when observation bias is present. </LI> </UL> </P> <P><B>Graphical abstract</B></P> <P>[DISPLAY OMISSION]</P>

Development of Gaussian Process-based Nonlinear Auto Regressive Exogenous (GP-NARX) surrogate model for dynamic analysis of natural gas liquefaction process

( Supaporn Wilailak ),이철진 한국공업화학회 2023 한국공업화학회 연구논문 초록집 Vol.2023 No.1

Drying characteristics of thermally pre-treated Cobra 26 F1 tomato slabs and applicability of Gaussian process regression-based models for the prediction of experimental kinetic data

Oladayo Adeyi,Emmanuel Olusola Oke,Abiola John Adeyi,Bernard Iberzim Okolo,Abayomi Olusegun Olalere,John Adebayo Otolorin,Ayomide Adeola,Brown Dagogo,Akinola David Ogunsola,Sunday Oladunni 한국화학공학회 2022 Korean Journal of Chemical Engineering Vol.39 No.5

The drying characteristics of unblanched (UB), steam blanched (SB) and hot water blanched (WB) Cobra26 F1 tomatoes were investigated at drying temperature of 40, 50, 60 and 70 oC and constant air velocity of 1.2m/s in aconvective oven. Gaussian process regression (GPR)-based models defined with squared-exponential kernel (GPR-SE),rational quadratic kernel (GPR-RQ), Matérn 5/2 kernel (GPR-M 5/2) and exponential kernel (GPR-Ex) were employedto model and predict experimental kinetic data of UB, SB and WB samples. Blanching and increased drying temperaturereduced the drying time. The effective moisture diffusivity, activation energy, total and specific energy requirementfor UB, SB and WB ranged between 3.6466 E -10 - 2.5526 E -09m2/s, 27.86-43.65 kJ/mol, 7.08-18.33kW-h and1,069.12-2,768.80kW-h/kg, respectively. Increased drying temperature and pre-treatment reduced activation energy, totaland specific energy requirements of Cobra 26 F1 tomatoes. Investigated GPR-based models were suitable for modellingand prediction of experimental kinetic data of Cobra 26 F1 tomatoes, GPR-M 5/2 was, however, marginally better. Hence, GPR-based models showed high suitability in handling multi-dimensional drying variables and can be used fordeveloping robust controllers applicable in auto-monitoring and control of Cobra 26 F1 tomatoes industrial drying.

기계학습 기법과 해색 위성 자료를 이용한 실시간 저층 용존산소농도산출기술 개발 연구

박성식,김경회 한국해양환경·에너지학회 2024 한국해양환경·에너지학회지 Vol.27 No.2

본 연구에서는 해색 위성 자료와 해양환경측정망 자료를 활용하여 시공간적 고해상도의 연안 저층 용존산소(dissolved oxygen, DO) 농도 산출을 위한 기계학습 모델을 개발하였다. 저층 DO 농도 산출을 위한 최적 모델로는 Gaussian process regression이 선별되었으며, 최적 예측변수로는 해색 위성 자료 중[원격반사도 6종, chlorophyll a 농도, 입자태 유기 탄소 농도, 분산광 소산 계수, 해수면 온도]가 선별되었다. 최적 예측변수 및 모델로 빈산소수괴가 가장 빈번하게 발생했던 대한해협의 저층 DO 농도를 산출했으며, 그 산출치와 관측치 간의 결정계수(R2)와 평균제곱오차(MSE)는 각각 0.69, 1.23으로 나타났다. 이후, 모델 정확도 개선 과정을 거친 최종 모델의 R2와 MSE는 각각 0.83, 0.47로 정확도 개선 전 대비 20.3, 61.8% 개선된 결과를 보였다. 매년 여름철 빈산소수괴로 어업 피해가 발생하고 있는 현재, 본 기술은 실시간 빈산소수괴 발생 탐지를 위한 기초 기술로 활용될 수 있을 것이다. In this study, we developed machine learning-based models for the spatiotemporal high-resolution estimation of coastal bottom dissolved oxygen (DO) concentration using ocean color data and marine environmental monitoring system data. Gaussian process regression was selected as the optimal model for bottom DO concentration estimation, and the optimal predictor variables were chosen as six types of remote sensing reflectance, chlorophyll-a, particulate organic carbon concentration, diffuse attenuation coefficient, and sea surface temperature. Utilizing the optimal predictor variables and model, we estimated the bottom DO concentration in the South Sea of Korea, which is prone to frequent hypoxia events. The coefficient of determination (R2) and mean squared error (MSE) between the estimated and observed values were 0.69 and 1.23, respectively. Subsequently, after a process of model accuracy improvement, the R2 and MSE of the final model were enhanced to 0.83 and 0.47, indicating a 20.3% and 61.8% improvement, respectively, compared to the accuracy before improvement. Given the recurring hypoxia events causing damage to fisheries, the current technology could be employed as a fundamental tool for real-time detection of hypoxia water mass, offering a promising solution.

Learning representative exemplars using one-class Gaussian process regression

Son, Youngdoo,Lee, Sujee,Park, Saerom,Lee, Jaewook Pergamon Press 2018 Pattern Recognition Vol. No.

<P><B>Abstract</B></P> <P>An exemplar is an observation that represents a group of similar observations. Exemplars from data are examined to divide entire heterogeneous data into several homogeneous subgroups, wherein each subgroup is represented by an exemplar. With its inherent sparsity, an exemplar-based learning model provides a parsimonious model to represent or cluster large-scale data. A novel exemplar learning method using one-class Gaussian process (GP) regression is proposed in this study. The proposed method constructs data distribution support from one-class GP regression using automatic relevance determination prior and heterogeneous GP noise. Exemplars that correspond to the basis vectors of the constructed support function are then automatically located during the training process. The proposed method is applied to various data sets to examine its operability, characteristics of data representation, and cluster analysis. The exemplars of some real data generated by the proposed method are also reported.</P> <P><B>Highlights</B></P> <P> <UL> <LI> A novel sparse Bayesian algorithm for learning exemplars is proposed. </LI> <LI> The proposed method automatically locates exemplars among similar observations. </LI> <LI> Applications to data representation and cluster analysis are provided. </LI> <LI> Theoretical generalization error bound for the method is provided. </LI> </UL> </P>