Abnormal Pattern Detection in Time Series Data via Artificial Immune System Model

Abstract

Today almost all anomaly detection systems (anti-virus solutions, intruder de- tection systems) are programmed to recognize known signatures of anomalies. But the di culty these systems are facing is that they are not capable of iden- tifying known attacks with unknown signatures and also totally unknown attacks. Natural immune system is best known anomaly detection system which is ca- pable of acting upon known and unknown attacks to the body. Arti cial immune system based anomaly detection systems are promising research area which ap- proximates the natural immune system model to build anomaly detection systems which are capable of overcome the above mentioned drawback present in tradi- tional anomaly detection systems. This project has been conducted with the aim of detecting anomalies appear- ing in daily transactions of stock exchange. The project approaches the problem by monitoring the daily price/volume data stream and the behavior of individu- als. Implemented solution bene ted by techniques used in natural immune system such as danger theory, negative selection, clonal selection and immune network theory which ultimately guide to identify unknown signatures of anomalies in daily transactions. It doesn't have a separate learning phrase but is capable of identifying the abnormalities by examining features of the given data stream rather than globally assigning boundary values to di erentiate normal and ab- normal ranges. System is tested on data collected from Colombo Stock Exchange and the re- sults were examined by a domain expert. Solution has been able to detect 26 abnormal movements out of 29 and captured all seven suspected behaviors by individuals. Four out of these seven scenarios identi ed as known signatures and others as unknown. System has identi ed another 9 price/volume uctuations as abnormal but they were not accepted.