Search

This paper presents a novel application of the theory of conformal prediction for distribution-independent on-line learning and anomaly detection. We exploit the fact that conformal predictors give valid prediction sets at specified confidence levels under the relatively weak assumption that the (normal) training data together with (normal) observations to be predicted have been generated from the same distribution. If the actual observation is not included in the possibly empty prediction set, it is classified as anomalous at the corresponding significance level. Interpreting the significance level as an upper bound of the probability that a normal observation is mistakenly classified as anomalous, we can conveniently adjust the sensitivity to anomalies while controlling the rate of false alarms without having to find any application specific thresholds. The proposed method has been evaluated in the domain of sea surveillance using recorded data assumed to be normal. The validity of the prediction sets is justified by the empirical error rate which is just below the significance level. In addition, experiments with simulated anomalous data indicate that anomaly detection sensitivity is superior to that of two previously proposed methods.

This paper presents a first attempt to evaluate two previously proposed methods for statistical anomaly detection in sea traffic, namely the Gaussian Mixture Model (GMM) and the adaptive Kernel Density Estimator (KDE). A novel performance measure related to anomaly detection, together with an intermediate performance measure related to normalcy modeling, are proposed and evaluated using recorded AIS data of vessel traffic and simulated anomalous trajectories. The normalcy modeling evaluation indicates that KDE more accurately captures finer details of normal data. Yet, results from anomaly detection show no significant difference between the two techniques and the performance of both is considered suboptimal. Part of the explanation is that the methods are based on a rather artificial division of data into geographical cells. The paper therefore discusses other clustering approaches based on more informed features of data and more background knowledge regarding the structure and natural classes of the data.