A study of Adversarial Machine Learning for Cybersecurity

Sam Afzal-Houshmand

Research output: Book/ReportPh.D. thesis

138 Downloads (Pure)

Abstract

The evolution of Internet of Things is expected to have a major impact on the lives of citizens as new services can be developed by the integration of the physical and digital worlds. However, with this increased richness there is also an emergence of malicious actors. Basically there are hackers that attack the data itself or the models. The models currently in use are heuristics which are susceptible to machine learning techniques. The theory is that to solve the issue of machine learning we use machine learning to defend ourselves i.e. fighting fire with fire. Considering that one of the core advantages is the unprecedented amount of data available for safety-critical decision making, we must be able to control those risks i.e. enabling security towards this direction is the investigation of advanced intrusion detection with the use of machine/deep-learning algorithms capable of achieving enhanced security awareness. The goal of this research project was to engage artificial intelligence and data science technologies towards developing a unified adversarial classification framework for identifying complex cybersecurity threats on the Internet and other cloud-based networking paradigms, all while taking into account uncertainty of data provenance, used for the classification and simultaneously handling the necessary belief inference and propagation modelling. Towards this direction, the use of machine learning (ML) and Deep learning (DL) has become omnipresent especially considering the advancements in computational efficiency and the maturity of large datasets available everywhere. Their predictions are used to make decisions about a number of critical applications including (amongst others) security for identifying complex cybersecurity threats in the context of the Internet (e.g., malicious domains) and other networking environments. Concretely, this project will be taking a discourse from the standpoint of malicious Domain Name Server (DNS) detection. DNS is a critical Internet service resolving IP addresses into hostnames that is crucial for the day-to-day operations of most safety critical systems. DNS, however, is susceptible to a wide range of attacks ranging from Domain Hijacking, DNS Flooding and Distributed Reflection Denial of Service (DRDoS) to Cache Poisoning, DNS Hijacking, DNS Spoofing, DNS Tunneling, etc. Leveraging machine- and deep-learning related concepts in the presence of these types of adversarial tactics towards enhanced classification, detection and security awareness is the core pillar of Adversarial Machine Learning - an emerging research area was extensively investigated in the context of this project towards understanding and improving the effectiveness of AI methods in the presence of advanced adversaries. Streamlined, this project used available collected data sets containing information about known malicious attacks (e.g. phishing, botnets, viruses, man-in-the-middle attack, spoofing etc.), in the context of malicious DNS servers (but not exclusively), towards developing generic classification models that will be able to detect and protect against such advanced adversaries. The outcomes were frameworks that addressed the central objective of enhancing security against advanced adversaries using ML and DL techniques which were capable of synthetically generate ontologies emulating DNS traffic based on a set of parameters that was determined from research. From this framework a couple of papers emerged which addressed how generating ontologies could be viable and the good results methods such as ML and DL learners applied to such data structure would denote. Furthermore, the viability of using ML and DL techniques in the realm of cybersecurity was investigated using other sources such as Mobile Crowd Sourcing data. This resulted in multiple papers that also take into account advanced techniques in the general area of Artificial Intelligence such as Explainable A.I. With this new understanding and the proofs provided from extensive experimentation there should now be a general approach and viability in applying and investigating ML based models in cybersecurity and comprehending their robustness against advanced adversarial learners.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages110
Publication statusPublished - 2023

Fingerprint

Dive into the research topics of 'A study of Adversarial Machine Learning for Cybersecurity'. Together they form a unique fingerprint.

Cite this