Model globoke nevronske mreže za zaznavanje omrežnih vdorov

BOVHA, JURE (author), Pustišek, Matevž (mentor)

Široka uporaba medsebojno povezanih računalniških sistemov je postala temelj izboljšave naših življenj, hkrati pa je izpostavila ranljivosti, katere je mogoče izkoristiti in presegajo človeški nadzor. Zaradi teh ranljivosti so varnostni sistemi na področju kibernetske varnosti bistveni, za varno komunikacijo med omrežji. Eden izmed mehanizmov, ki nam omogoča obrambo pred omrežnimi napadi, so sistemi za zaznavanje vdorov IDS (angl. Intrusion Detection System). Glede na pričakovano pospešitev in povečanje računalniških groženj v tej nalogi raziskujem uporabnost in zmogljivost algoritmov globokega učenja na področju zaznavanja vdorov v omrežje. Naloga opisuje implementacijo modela globokega učenja, ki je naučen na naboru podatkov CIC-IDS-2018, dostopnem na spletni strani Kaggle. Ustvarjena arhitektura globoke nevronske mreže DNN (angl. Deep neural network) je ovrednotena z različnimi metrikami, bodisi s poudarkom na hitrosti ali pa na natančnosti. Nabor podatkov je predhodno obdelan, da lahko sistem za zaznavanje vdorov deluje z manj računskimi stroški. Testiranje rezultatov je ločeno glede na dejstvo, ali je opazovan segment omrežnega prometa del napada ali pa ne.

• Language: Slovenian
• Keywords: strojno učenje, klasifikacija, omrežni napadi
• Work type: Bachelor thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2023
• PID: 20.500.12556/RUL-144643
• COBISS.SI-ID: 144320771
• Publication date in RUL: 06.03.2023

Secondary language

• Language: English
• Title: A deep neural network model for network intrusion detection

Abstract

The widespread use of interconnected computer systems has become the foundation of improving our lives. At the same time it highlighted the vulnerabilities, that can be exploited to overcome human control. Because of these vulnerabilities, security systems in the field of cyber security are essential, for secure communication between networks. One of the mechanisms that allow us to defend against network attacks are intrusion detection systems (IDS). In this paper I explore the applicability and performance of deep learning algorithms in the field of knowledge of network intrusions, given the expected acceleration and increase in computer threats. The paper describes the implementation of a deep learning model, trained on the CIC-IDS-2018 dataset, which is available on the Kaggle website. The created architecture of the deep neural network (DNN) is evaluated with various metrics, with an emphasis on speed, or accuracy. The dataset is pre-processed, so that the intrusion detection system can operate with less computational cost. The testing of results is divided, so that the result either points to the network flow being part of an attack, or not.

• Keywords: machine learning, classification, network attacks