INVESTIGATIONS OF AUTOENCODER HYPER-PARAMETERS ON ANOMALY DETECTION

Abstract

Most of anomaly detection techniques, such as density-based methods, often perform inefficiently on the high dimension of network data because the curse of dimensionality phenomenon. Our previous work presented a novel approach of using the feature space of AutoEncoders (AEs) as a new feature representation for addressing this problem. In this study, we attempt to investigate the characteristics of the latent representation of AEs. Thus, we first discuss the hypothesis of using the latent representation in more details, and extend several experiments showed in the previous work. Following this, we design three intensive examinations (an investigation on the middle hidden layer size, an evaluation on the performance of the hybrid and an exploration on latent data). These aim to get insight into the latent representations of AEs, which is fundamental for designing good latent representations in the future work. This paper closes with analysis and discussion on the experimental results.