Anomalioiden havaitseminen langattomissa sensoriverkoissa syväoppimisen avulla
Abstract
Globaalista IP-verkkoliikenteestä yhä suuremmasta osuudesta vastuussa olevat uuden sukupolven langattomat verkot ja mobiili- sekä IoT-sovellukset ovat jalkautumassa aina kriittisen infrastruktuurin järjestelmiin asti. Fyysisen ja digitaalisen maailman rajapinnassa osana IoT-sovelluksia toimivat langattomat sensoriverkot ovat alttiita laajalle kirjolle erilaisia tietoturvauhkia niiden avoimen luonteen, IoT-sovellusten teknologisen kypsymättömyyden ja alati kehittyvän kyberrikollisuuden vuoksi. Langattomien sensoriverkkojen suojaaminen kyberhyökkäyksiltä ja muulta niiden luotettavaa toimintakykyä uhkaavalta ja vahingoittavalta toiminnalta on tärkeä tutkimusaihe. Tässä työssä tutkittiin hiljattain julkaistun esineiden internetin sovellusympäristöä jäljittelevän Bot-IoT -datajoukon avulla verkkohyökkäyksien tunnistamista anomalioiden havaitsemisen keinoin käyttämällä moderneja syväoppimismenetelmiä. Työssä implementoidaan ja vertaillaan neljää autoenkooderiarkkitehtuuriin perustuvaa yksinkertaista ja laskennallisesti kevyttä syväoppimismallia. Suorituskykyisin toistuvaan neuroverkkoon perustuva LSTM-autoenkooderi kykeni tunnistamaan yli 3,6 miljoonaa hyökkäystä jättäen vain 101 hyökkäystä tunnistamatta. Työssä tehdyn kaltaista tutkimusta Bot-IoT -datajoukkoon ei ole tiedeyhteisössä aiemmin toteutettu eikä vastaavia tuloksia ole ennen saatu. Lisäksi työssä annetaan kattava teoreettinen tausta tunnetuimmista syväoppimismenetelmistä ja niiden soveltamisesta anomalioiden havaitsemiseen.
The next-generation wireless and mobile networking as well as IoT applications accounting for an ever-increasing share of the global IP network traffic are being widely deployed reaching critical infrastructures. Acting as an interface between the physical and the digital world in IoT applications, wireless sensor networks are exposed to a wide range of information security threats due to their open nature of communications, the technological immaturity of IoT solutions and the accelerating growth of cybercrime. Protecting wireless sensor networks from cyberattacks and other factors that may impair the continuity of their secure and reliable operations is an important area of research. In this thesis, the ability of detecting network attacks with methods based on deep learning using principles from anomaly detection was investigated by a recently published dataset called Bot-IoT that incorporates flow-based network traffic from an IoT environment. Four different lightweight deep learning based autoencoders were implemented for evaluation and comparison purposes. The results demonstrated the superiority of the recurrent LSTM-autoencoder model by detecting over 3.6 million attacks while leaving only 101 attacks undetected. The empirical study conducted in this thesis with the Bot-IoT -dataset is the first of its kind in the scientific community and similar results have not yet been published. In addition, a comprehensive theoretical background of the most common deep learning methods and their applicability to anomaly detection is given.
The next-generation wireless and mobile networking as well as IoT applications accounting for an ever-increasing share of the global IP network traffic are being widely deployed reaching critical infrastructures. Acting as an interface between the physical and the digital world in IoT applications, wireless sensor networks are exposed to a wide range of information security threats due to their open nature of communications, the technological immaturity of IoT solutions and the accelerating growth of cybercrime. Protecting wireless sensor networks from cyberattacks and other factors that may impair the continuity of their secure and reliable operations is an important area of research. In this thesis, the ability of detecting network attacks with methods based on deep learning using principles from anomaly detection was investigated by a recently published dataset called Bot-IoT that incorporates flow-based network traffic from an IoT environment. Four different lightweight deep learning based autoencoders were implemented for evaluation and comparison purposes. The results demonstrated the superiority of the recurrent LSTM-autoencoder model by detecting over 3.6 million attacks while leaving only 101 attacks undetected. The empirical study conducted in this thesis with the Bot-IoT -dataset is the first of its kind in the scientific community and similar results have not yet been published. In addition, a comprehensive theoretical background of the most common deep learning methods and their applicability to anomaly detection is given.
Main Author
Format
Theses
Master thesis
Published
2019
Subjects
The permanent address of the publication
https://urn.fi/URN:NBN:fi:jyu-201912135267Use this for linking
Language
Finnish