近年，深層学習を代表とする機械学習が脚光を浴びており，様々なシステムへの応用が検討されている．応用対象は，エンターテイメントシステムや事務処理システムなどにとどまらず，自動車の自動運転などの高安全システムなども検討されている．高安全システムでは，安全性に関する十分なテストを実施することが必要である．一方，機械学習システムの安全性をテストする手法は，まだ確立していない．そこで，本論文では，機械学習を用いて実現された分類器に焦点を当て，その安全性を系統的に評価する手法を提案する．提案手法は，データセットに基づいた安全分析とテスト結果の統計的評価方法から構成される．安全分析では，よく知られている FTA (Fault Tree Analysis) をデータセットを取り扱えるように拡張した．そして，安全分析の結果，期待される認識率の水準を求め，テストを実施する．テストでは，この水準を満たしているか統計的に評価を実施する．提案手法は，CNN (Convolutional Neural Network) による手書き文字の分類器に適用し，評価を行った．

Recently, machine learning, especially, deep learning, accumulates interests and applying it to various kinds of systems is being studied. The targets of applications include not only entertainment systems but also safety critical systems such as autonomous driving of car. In a safety critical system, it is indispensable to ensure the safety of the system before using and deploying it in our society. However, how we ensure the safety of systems which use machine learning has not been established yet. In this paper, we propose a method to systematically evaluate the safety of such systems. The proposed method consists of the safety analysis based on datasets and the statistical evaluation of testing results. In the safety analysis, widely known and used FTA (Fault Tree Analysis) is extended to deal with the datasets and it allows us to obtain expected recognition rates of the systems. Then, we conduct the testing of systems in order to check whether it is met or not. The satisfaction of the expected recognition rates is ensured by statistic evaluation of the testing result. We also conducted experiments using a handwritten character recognition system which is realized by CNN (Convolutional Neural Network) to show the feasibility and effectiveness of our method. We focus on supervised classifiers realized by machine learning in this paper but we believe that it can be extended to the other kinds of machine learning applications.