Artificial Immune Systems for Self-Nonself Discrimination: Application to Anomaly Detection
Self-nonself discrimination is the ability of the vertebrate immune systems to distinguish between foreign objects and the body’s own self. It provides the basis for several biologically inspired approachs for classification. The negative selection algori
- PDF / 4,546,758 Bytes
- 18 Pages / 439.2 x 666 pts Page_size
- 82 Downloads / 160 Views
uction One of the fundamental characteristics of the mammalian immune system is the ability to recognize the presence of foreign bodies called pathogens [1]. This process is carried out by means of antibodies, which are receptor sites attached to lymphocytes and capable of binding to pathogens. There are two kinds of antibody producing cells, the T-cells and the B-cells. T-cell antibodies possess the ability to distinguish invading antigens from the body’s own cells. This phenomenon is termed as self-nonself discrimination. Self-nonself discrimination proceeds within a specimen’s immune system by generating a large variety of detector cells, systematically culling out those that erroneously categorize native cells (self) as foreign (nonself) while retaining the rest. This principle is called negative selection [2]. Self-nonself discrimination is becoming increasingly popular as a paradigm in industrial and other applications. An algorithm based on this approach typically works in the following manner. During the training phase, a large amount S. Das et al.: Artificial Immune Systems for Self-Nonself Discrimination: Application to Anomaly Detection, Studies in Computational Intelligence (SCI) 116, 231–248 (2008) c Springer-Verlag Berlin Heidelberg 2008 www.springerlink.com
232
S. Das et al.
of samples of normal data are presented to the algorithm, allowing it to adapt. These samples are referred to as self samples. In the input space, the region containing all possible self samples is called the self region, or simply self (S). Its complement is the nonself region, or nonself. (In an example that we have considered later in this chapter, the entire self space will be defined a priori, instead of through self samples.) A fully trained algorithm, when presented any sample from the input space should be capable of correctly classifying it as either self or nonself. This feat is not readily accomplished with a conventional machine learning classification scheme as proper training requires the availability of enough samples of both classes, self as well as nonself. Data pertaining to nonself is usually sparse in most industrial applications, where nonself is linked to abnormal conditions. Negative selection algorithms use a set, D, of detectors, which also correspond to points in the input space. These detectors are the biological analogues of the B-cell antibodies. During the training phase, a sufficient number of detectors are generated to fill the detector set. In each iteration, a candidate detector is generated, usually at random. It is then cross-checked with the self sample that are available for training to see if it can detect any such point. If it does, the detector is discarded, otherwise it is inserted into the detector set. Training is terminated only after a large number of detectors are found. During the detection phase, the input is checked to see if it can be detected by any existing detector. When it does, the input is classified as a nonself sample. On the other hand, an input that cannot be detected is l
Data Loading...
