Estimation of applicability of modern neural network methods for preventing cyberthreats to self-organizing network infrastructures of digital economy platforms,

Maxim Kalinin,Vasiliy Krundyshev,Evgeny Zubkov,A Sarygulov,W Semmler,L Ungvári,V Sergeev

doi:10.1051/shsconf/20184400044

Maxim Kalinin, Vasiliy Krundyshev + Show 5 more

Open Access

https://doi.org/10.1051/shsconf/20184400044

Copy DOI

Abstract

The problems of applying neural network methods for solving problems of preventing cyberthreats to flexible self-organizing network infrastructures of digital economy platforms: vehicle adhoc networks, wireless sensor networks, industrial IoT, “smart buildings” and “smart cities” are considered. The applicability of the classic perceptron neural network, recurrent, deep, LSTM neural networks and neural networks ensembles in the restricting conditions of fast training and big data processing are estimated. The use of neural networks with a complex architecture– recurrent and LSTM neural networks – is experimentally justified for building a system of intrusion detection for self-organizing network infrastructures.

Highlights

Self-organizing network infrastructures have peer-to-peer architecture, support multiple links between hosts and dynamic control of routing at each network node that determines their advantages over traditional networks, that is the possibility of multi-line transfer of data to large distances without stationary retranslators, immunity to spacial variations in network topology, dynamic reconfiguration of a network under the conditions of distorting actions and faults [1]
The basis for solving the specified problem is made of different types of artificial neural networks (ANN) which have appeared last decade; including deep ANN, recurrent networks, LSTM networks, and their ensembles
Comparing the characteristics of a recurrent ANN and a LSTM network (Fig. 2) demonstrates their similarity, a recurrent ANN has a series of drops in the process of training which are missing in the analogous process in an ANN with memory because the latter uses the optimization mechanism – the memory enables locking the result through a feed-back and remembering the conclusions made

Summary

Introduction

As introduction of security means into inter-computer network is a complicated procedure due to the peculiarities of networks of such a type (decentralization, dynamics, peer-to-peer character, etc.) and the limitations of computational resources on hosts, active demand for methods of aprior protection is formed These methods should provide for timely determining of cyberthreats and, increasing level of protection without interfering network infrastructure. The classic perceptron of direct signal distribution and actively developing ANNs have been selected for estimation of applicability and correlation of modern ANNs in comparison with traditional ones for solving the problem of detecting cyberthreats aimed at flexible network infrastructures: recurrent ANN [9], deep ANN [10], LSTM neural network [11], and ensemble of neural networks [12]. The results obtained in the course of training the investigated ANNs are listed in table 1

Deep ANN

Findings

Conclusion