2026-04-12T19:39:32Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00197603 2025-01-19T22:17:19Z 1164:5352:9740:9828

敵対的外乱とTriple GANsとの組み合わせによる識別的異常検知 Triple GANs with adversarial disturbances for discriminative anomaly detection 八谷, 大岳 Hirotaka, Hachiya Anomaly detection (AD) is an important machine learning task to detect outliers given only normal training data—applied to real-world problems such as surveillance camera. Recently, an advanced deep learning technique, generative adversarial nets (GANs) has been applied to AD, so that the generator is used to generate virtual anomalies and the discriminator is trained to classify instances into normal or (virtual) abnormal classes. However, existing GANs approach would have two problems: 1) its performance is highly dependant on hyper parameters on early-stopping to build intentionally an imperfect generator, 2) the detector tends to excessively focus on specific features, called single mode. In this research, to overcome these problems, we propose to introduce an additional discriminator to GANs, specifically for the anomaly detection, i.e., triple GANs, and then train it with virtual anomalies generated by the generator with latent adversarial disturbances. We show the effectiveness of our proposed method, called. Triple GANomalies, through toy-data experiments using MNIST dataset. Anomaly detection (AD) is an important machine learning task to detect outliers given only normal training data—applied to real-world problems such as surveillance camera. Recently, an advanced deep learning technique, generative adversarial nets (GANs) has been applied to AD, so that the generator is used to generate virtual anomalies and the discriminator is trained to classify instances into normal or (virtual) abnormal classes. However, existing GANs approach would have two problems: 1) its performance is highly dependant on hyper parameters on early-stopping to build intentionally an imperfect generator, 2) the detector tends to excessively focus on specific features, called single mode. In this research, to overcome these problems, we propose to introduce an additional discriminator to GANs, specifically for the anomaly detection, i.e., triple GANs, and then train it with virtual anomalies generated by the generator with latent adversarial disturbances. We show the effectiveness of our proposed method, called. Triple GANomalies, through toy-data experiments using MNIST dataset. 情報処理学会 2019-06-10 eng technical report https://ipsj.ixsq.nii.ac.jp/records/197603 2188-8590 AA12055912 研究報告バイオ情報学（BIO） 2019-BIO-58 4 1 6 https://ipsj.ixsq.nii.ac.jp/record/197603/files/IPSJ-BIO19058004.pdf application/pdf 475.8 kB