샴 신경망 출력값의 우도비 기반 동일인 필적 여부에 관한 분석 (Case Analyses of Identical Handwriting Based on Likelihood Ratio of Siamese Neural Network Output)

초록

필적정보를통한생체인식은유서, 계약서등의진위여부판별을통해중요한법적증거가될수있 다. 그러나 일반적으로 필기 쌍의 동일인 여부는 전문가의 경험과 주관에 의해 결정된다. 인공지능을 통한 생체인 식이법정증거력을확보하기위해서는판정에대한신뢰도값이제시될필요가있다. 확률분포및최소오류결 정이론기반의이진분류문제에서는판정에대한신뢰도를우도비로산출할수있다. 본논문에서는샴신경망 기반필적생체정보의동일인및타인매칭점수분포를확률분포형태로구축하고, 비교되는필적쌍의점수를 통해 본인 여부 및 우도비 값을 산출한다. 동일인 분포와 타인 분포가 겹쳐지는 구간에서는 판정에 대한 우도비가 급격하게 작아진다. 실제로 우도비가 작게 산출되는 경우가 인간의 육안으로도 판정의 모호성을 가지는지 정성적 으로 분석해본다. 분석 결과, 우도비가 큰 경우에는 육안으로도 판정이 명확한 경우가 대부분이었으며, 우도비가 작은 경우에는 실제로 육안 판정이 어려운 것으로 나타났다.

영어초록

Handwriting biometrics can be important legal evidence through determination of the authenticity of wills and contracts. However, in general, whether handwriting pairs are genuinity is determined by the expert's experience and subjectivity. In order for biometric recognition through artificial intelligence to secure legal evidence, it is necessary to present a reliability value for the genuinity decision. In binary classification problem based on probability distribution and least-error decision theory, the reliability of the decision can be calculated as a likelihood ratio. In this paper, genuine and imposter matching score distributions of handwriting biometrics based on Siamese neural network is constructed in the form of probability distribution, and the identity and likelihood ratio values a​re calculated through the scores of handwriting pairs compared. In the ovelapped section between the genuine and the imposter distributions , the likelihood ratio for the judgment decreases sharply. Qualitatively analyze whether the case in which the likelihood ratio is actually calculated has ambiguity in the judgment even with the human eye. As a result of the analysis, it was found that, in most cases, when the likelihood ratio was large, the determination was clear even with the naked eye, and when the likelihood ratio was small, it was actually difficult to determine visually.

참고자료

· 없음