Development and Research of an Evolutionary Algorithm for the Formation of a Feature Space based on AutoML for Solving the Problem of Identifying Cyber Attacks
| Autor: | Lubov Zabrodina, Irina Bolodurina, Denis Parfenov, Arthur Zhigalov, Alexander Shukhman |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
0106 biological sciences
business.industry Computer science Feature vector Evolutionary algorithm Machine learning computer.software_genre 010603 evolutionary biology 01 natural sciences Evolutionary computation 010602 entomology Tree (data structure) Statistical classification Feature (computer vision) Artificial intelligence business computer Feature learning Selection (genetic algorithm) |
| Zdroj: | 2020 International Conference Engineering and Telecommunication (En&T). |
| DOI: | 10.1109/ent50437.2020.9431251 |
| Popis: | Currently, the use of intelligent data analysis methods allows us to solve applied problems that go beyond simple machine learning and determines the development vector for the possibility to automate all stages of this process. This article presents an approach that allows you to automate the construction and selection of features of raw datasets. The developed method is a logical extension of the ExploreKit algorithm at the stage of generating new feature candidates and calculating meta-objects to represent the dataset and feature candidates. Also, our approach integrated the tree representation of transformations of the AutoFE feature space and methods of evolutionary optimization. Experimental studies on UNSW-NB15, CICDDoS2019, and APA-DDoS traffic datasets have shown that the evolutionary algorithm to form the feature space allows obtaining the necessary accuracy comparable to other algorithms. In addition, the evolutionary approach showed good performance in most cases due to the parallelization of calculations. On the CICDDoS dataset with the highest number of characteristics, the performance of evolutionary approach was 7.4% higher than the Feature Learning by Tree algorithm, but our approach was 13.9% lower than the Feature Learning by Deep Network algorithm. However, the accuracy of FDN algorithm was 2.2 % lower than the accuracy of the evolutionary approach. On the UNSW_NB15 and APA-DDoS datasets, the developed approach showed a decrease in performance, but its accuracy compared to other algorithms is 1.9% higher on average. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|