Adaptive Approach for Density-Approximating Neural Network Models for Anomaly Detection

We propose an alternative use of neural models in anomaly detection. Traditionally, in anomaly detection context the common use of neural models is in form of auto-encoders. Through the use of auto-encoders the true anomality is proxied by reconstruction error. Auto-encoders often perform well but do not guarantee to perform as expected in all cases. A popular more direct way of modeling anomality distribution is through k-Nearest Neighbor models. Although kNN can perform better than auto-encoders in some cases, their applicability can be seriously impaired by their space and time complexity especially with high-dimensional large-scale data. The alternative we propose is to model the distribution imposed by kNN using neural networks. We show that such neural models are capable of achieving comparable accuracy to kNN while reducing computational complexity by orders of magnitude. The de-noising effect of a neural model with limited number of neurons and layers is shown to lead to accuracy improvements in some cases. We evaluate the proposed idea against standard kNN and auto-encoders on a large set of benchmark data and show that in majority of cases it is possible to improve on accuracy or computational cost.