ScholarWorks@숙명여자대학교

초록

Anomaly detection is to detect data that deviate significantly from the pattern of the majority of data. The autoencoder is a representative model used for anomaly detection. An autoencoder first learns the features of normal data, and then determines data that greatly deviate from the the learned features as anomalous data. In order to learn the features of normal data, an autoencoder needs a large amount of data labeled as normal. However, it is usually not easy to obtain a large amount of labeled data. To address this problem, this paper proposes a method that utilizes even unlabeled data to train an autoencoder, which is based on the meta pseudo labeling. The proposed method selects data that is predicted as normal from unlabeled data and additionally uses them to train an autoencoder. Consequently, the proposed method can improve the performance of anomaly detection by using unlabeled data even in an environment where there are not much labeled data. Through experiments using two real datasets, we confirmed that the proposed method detects anomalies more accurately than the existing anomaly detection techniques.

키워드