Research On Detector Generation And Self Representation Methods

The Negative Selection Algorithms (NSAs) are a kind of Artificial Immune Algorithms, which are inspired by the self-nonself discrimination in the nature immune system. The Real-valued Negative Selection Algorithms (RNSAs) are an important branch of the NSAs. They adopt a real-valued representation of the space to solve the anomaly detection problems in real-valued space.In a NSA, a set of self samples must be defined before the algorithm. The self samples represent the normal states of a system. Due to the finiteness of the self samples, the self samples should be generalized by some methods to cover as much self space as possible. But at the same time, it should not cover much nonself space. Then the mature detectors, which do not match with any of the self samples, are generated. Finally, the mature detector set is used to detect the anomalous behaviors or states of a system.This thesis aims at the detector generation and the self representation of the NSAs. For the two important problems of the NSAs, a novel detector generation algorithm and a novel self representation method are proposed. The specific works include the following aspects.(1) The random detector generation algorithm had a low efficiency to generate an optimal detector set. In this thesis, a detector evolutionary generation algorithm, named as EvoSeedRNSA, is proposed. The EvoSeedRNSA uses an evolutionary algorithm to evolve the random seeds to generate an approximately optimal detector set. The experimental results of the comparisons demonstrated that the EvoSeedRNSA gained a great increment of the detection rates.(2) On the basis of the EvoSeedRNSA, an improved detector evolutionary generation algorithm, named as EvoSeedRNSAⅡ, is proposed. By redesigning of the individual encoding and the genetic operators, the better schemas in an individual can be more effectively preserved by the EvoSeedRNSAⅡ. The experimental results demonstrate that, as the increment of the number of random seeds that included in an individual, the detection rates of the EvoSeedRNSAⅡhave been significantly increased.(3) In order to deal with the deficiencies of the traditionally self representation based on fixed self radius (FSR), this thesis proposes a novel self representation based on an adaptive self radius. Firstly, a self representation K-NN-ASR, based on the K nearest neighbor algorithm (K-NN), is given. Then a novel boundary judgment method is introduced. Combined with the boundary judgment method and the K-NN algorithm, a self representation B-NN-ASR is proposed. In the experiments, the performance of the k-NN-ASR, B-NN-ASR and FSR are compared. The experimental results demonstrate that the B-NN-ASR has a higher self coverage and lower nonself coverage.In summary, this thesis proposes a novel detector evolutionary generation algorithm and a novel self representation method. Good results are achieved in both aspects of the NSAs, i.e. the detector generation and the self representation.