Online Reactive Anomaly Detection over Stream Data

Outlier detection over data streams has attracted attention for many emerging applications, such as network intrusion detection, web click stream and aircraft health anomaly detection. Since the data stream is likely to change over time, it is important to be able to modify the outlier detection model appropriately with the evolution of the stream. Most existing approaches were using incremental or periodical models to deal with evolving stream data. However, in these approaches, model updates were either more frequently and risk wasting resources on insignificant changes or more infrequently and risk model inaccuracy. In this paper, a hybrid framework by combining LOF (local outlier factor) and BPNN (back propagation neural network), appropriate for online detecting outliers in data streams, is proposed. The proposed framework provides equivalent detection performance as the iterated static LOF algorithm (applied after insertion of each data record), while requiring significantly less computational time.