Optimization of Real-Valued Self Set for Anomaly Detection Using Gaussian Distribution

The real-valued negative selection algorithm (RNS) has been a key algorithm of anomaly detection. However, the self set which is used to train detectors has some problems, such as the wrong samples; boundary invasion and the overlapping among the self samples. Due to the fact that the probability of most real-valued self vectors is near to Gaussian distribution, this paper proposes a new method which uses Gaussian distribution theory to optimize the self set before training stage. The method was tested by 2-dimensional synthetic data and real network data. Experimental results show that, the new method effectively solves the problems mentioned before.