Anomaly detection by auto-association

Anomaly detectors (or novelty detectors) are systems for detecting behaviour that deviates from "normality ", and are useful in a wide range of surveillance, monitoring and diagnosis applications. Feed-forward auto-associative neural networks have, in several studies, shown to be effective anomaly detectors although they have a tendency to produce false negatives. Existing methods rely on anomalous examples (counter-examples) during training to prevent this problem. However, counter-examples may be hard to obtain in practical anomaly detection scenarios. We therefore propose a training scheme based on regularisation, which both reduces the problem of false negatives and also speeds up the training process, without relying on counter-examples. Experimental results on benchmark machine learning problems verify the potential of the proposed approach