Earticle

현대 신경망은 종종 높은 분류 정확도를 달성하지만, 여전히 신뢰도 보정 오류 문제를 겪고 있어 의사결정이 중요한 응용 분야에서 신뢰성이 제한된다. 본 논문에서는 신경망 신뢰도의 신뢰성을 향상시키기 위해 MADCAT이라 불리는 2단계 사후 보정 방법을 제안한다. 제안된 방법은 먼저 로짓 마진을 샘플 난이도의 지표로 사용하여 샘플별 온도를 예측하고, 차선 레이어 특징 공간에서 K-최근접 이웃 거리를 기반으로 한 밀도 인식 조정을 통해 보정된 신뢰도를 정교화 한다. 또한, 최적화와 보정 품질 간의 정합성을 높이기 위해 훈련 목표로 소프트-빈 기대 보정 오차를 채택한다. CIFAR-10, CIFAR-100 및 ImageNet에 대한 실험결과, 제안된 방법은 컨볼루션 신경망과 비전 트랜스포머 모두에서 뛰어난 보정 성능을 달성하는 동시에, 클래스 불균형, 분포 이동 및 제한된 검증 데이터 환경에서도 미미한 추론 오버헤드로 효과적인 성능을 유지하는 것으로 나타났다.

Modern neural networks often achieve high classification accuracy but still suffer from confidence miscalibration, which limits their reliability in decision-sensitive applications. In this paper, a two stage post-hoc calibration method, called MADCAT, is proposed to improve the reliability of neural network confidence. The proposed method first uses the logit margin as an indicator of sample difficulty to predict a sample-specific temperature, and then refines the calibrated confidence through a density-aware adjustment based on K-Nearest Neighbor distances in the penultimate-layer feature space. In addition, Soft-Binned Expected Calibration Error is adopted as the training objective to better align optimization with calibration quality. Experimental results on CIFAR-10, CIFAR-100, and ImageNet show that the proposed method achieves strong calibration performance on both convolutional neural networks and Vision Transformers, while remaining effective under class imbalance, distribution shift, and limited validation data settings with negligible inference overhead.

This paper presents a margin-based face liveness detection method with behavioral confirmation to prevent spoofing attacks using deep learning techniques. The proposed method provides a possibility to prevent biometric person authentication systems from replay and printed spoofing attacks. For this work, a set of real face images and fake face images was collected and a face liveness detection model is trained on the constructed dataset. Traditional face liveness detection methods exploit the face image covering only the face regions of the human head image. However, outside of this region of interest (ROI) might include useful features such as phone edges and fingers. The proposed face liveness detection method was experimentally tested on the author’s own dataset. Collected databases are trained and experimental results show that the trained model distinguishes real face images and fake images correctly.

This paper describes a reactive optimal power flow incorporating margin enhancement constraints. Margin sensitivity at a steady-state voltage instability point is calculated using invariant space parametric sensitivity, and it can provide valuable information for selection of effective control parameters. However, the weakest buses in neighboring regions have high margin sensitivities within a certain range. Hence, the control determination using only the sensitivity information might cause violation of operational limits of the base operating point, at which the control is applied to enhance voltage stability margin in the direction of parameter increase. This paper applies an interior point method (IPM) to solve the optimal power flow formulation with the margin enhancement constraints, and shunt capacitances are mainly considered as control variables. In addition, nonlinearity of margin enhancement with respect to control of shunt capacitance is considered for speed-up control determination in the numerical example using the IEEE 118-bus test system.

보험의 위험률은 통계를 바탕으로 생성되며, 미래에 대한 불확실성으로 위험률 산출 시 리스크에 대한 대비책으로 안전할증이라는 것을 설정한다. 현재 보험 산업에서는 일반적으로 통계적 표준편차를 이용하여 안전할증을 산출하고 있으나 발생률이 증가 추세를 가질 경우 리스크에 대한 대비책으로 충분치 못한 경우가 있다. 따라서, 본 연구에서는 위험률이 가지는 리스크를 수준리스크, 일반추세리스크, 변동성리스크, 재난리스크로 분류하여 정의하고 제반 특징을 살펴보았다. 그리고 이 중 수준리스크와 일반추세리스크의 크기를 측정하는 방법과 이를 바탕으로 한 안전할증을 산출하는 방법을 제시하고, 통계청 및 국립암센터의 자료를 이용한 암발생률에 적용하여 안전할증을 산출하여 보았다. 위험률이 가지는 리스크를 분석 및 측정하여 이에 대응하는 안전할증을 산출하였기 때문에 안정적인 위험률관리 측면에서 기존의 안전할증보다 우수함을 알 수 있다.국문색인어 : 위험률, 리스크, 안전할증

Mortality and morbidity rates on which insurance is based are created by statistics. Safety margin is added to prevent risk because of future uncertainty when mortality and morbidity rate are created. Safety margin is calculated by statistical method using standard deviations in current insurance industry. However, safety margin may not be enough for risk caused from incident rate's trend.So, in this study we classify risks that incident rate can have into 4 cases, level risk, general trend risk, volatility risk, and calamity risk. We show how to calculate level risk and trend risk and how to calculate safety margin based on level risk and general trend risk. We give a concert example with cancer incident rate.Since we make safety margin based on analysis and measurement of risks of mortality and morbidity rates, our safety margin is better than current safety margin in mortality and morbidity rates management.