[2I4-GS-2] Machine learning: Anomaly detection

[2I4-GS-2-02] An Anomaly Detection Method with Neural Network Near Neighbor

〇Yuichi Kato¹, Kentaro Takagi¹, yaling Tao¹, Susumu Naito¹, Yasunori Taguchi¹, Teguh Budianto¹, Kouta Nakata¹ (1. TOSHIBA CORPORATION)

Keywords:Neural Network, k-nearest neighbor, anomaly detection

Machine learning for anomaly detection is a key technology employed in diverse industries. k-nearest neighbor and AutoEncoder are one of those methods used frequently. However, the former one searches nearest neighbors even for anomalies requiring large distance between normal and anomaly examples, and the latter one suffers from excessive generalization. In order to tackle the problems, we developed an algorithm, neural network near neighbor, approximating neighborhood search by a neural network. It seeks out train data addition ratio to reproduce input data via a softmax layer. Therefore, it neither derives nearest neigbors of anomaly examples nor recreates data which are unattainable from superposition of train data. We evaluated the performance of it with MNIST data-set consisting of handwritten digits. The algorithm has the highest mean of the area under the curve of Reciever Operating Characteristic of 0.850, while k-nearest neighbor and AutoEncoder show 0.822 and 0.623, respectively.