Herein, we study an insider anomaly behavior detection method using time-series-based log data that records insider behavior. We developed a long-short-term-memory-based denoised autoencoder model and extracted latent vectors containing useful sequenc...
Herein, we study an insider anomaly behavior detection method using time-series-based log data that records insider behavior. We developed a long-short-term-memory-based denoised autoencoder model and extracted latent vectors containing useful sequence information from the autoencoder. The performance of the insider anomaly detection method was further evaluated by inputting the extracted latent vectors to anomaly detection algorithms—Local Outlier Factor and Isolation Forest. By verifying the effectiveness of the model using various performance evaluation indicators, via the coding vector (dimension: 5), it was confirmed that the shorter the sequence length, the higher the recall, and using the coding vector (dimension: 7), the higher the recall regardless of the sequence length. Furthermore, while keeping the number of abnormal behavior samples constant, it was confirmed that the precision decreased as the number of normal behavior samples increased.