Spectrum Sensing Security in Cognitive Radio Networks

Loading...
Thumbnail Image

Files

Publication or External Link

Date

2010

Citation

DRUM DOI

Abstract

This thesis explores the use of unsupervised machine learning for spectrum sensing in cognitive radio (CR) networks from a security perspective. CR is an enabling technology for dynamic spectrum access (DSA) because of a CR's ability to reconfigure itself in a smart way. CR can adapt and use unoccupied spectrum with the help of spectrum sensing and DSA. DSA is an efficient way to dynamically allocate white spaces (unutilized spectrum) to other CR users in order to tackle the spectrum scarcity problem and improve spectral efficiency. So far various techniques have been developed to efficiently detect and

classify signals in a DSA environment. Neural network techniques, especially those using unsupervised learning have some key advantages over other methods mainly because of the fact that minimal preconfiguration is required to sense the spectrum. However, recent results have shown some possible security vulnerabilities, which can be exploited by adversarial users to gain unrestricted access to spectrum by fooling signal classifiers. It is very important to address these new classes of security threats and challenges in order to make CR a long-term commercially viable concept.

This thesis identifies some key security vulnerabilities when unsupervised machine learning is used for spectrum sensing and also proposes mitigation techniques to counter the security threats. The simulation work demonstrates the ability of malicious user to manipulate signals in such a way to confuse signal classifier. The signal classifier is forced by the malicious user to draw incorrect decision boundaries by presenting signal features which are akin to a primary user. Hence, a malicious user is able to classify itself as a primary user and thus gains unrivaled access to the spectrum. First, performance of various classification algorithms are evaluated. K-means and weighted classification algorithms are selected because of their robustness against proposed attacks as compared to other classification algorithm. Second, connection attack, point cluster attack, and random noise attack are shown to have an adverse effect on classification algorithms. In the end, some mitigation techniques are proposed to counter the effect of these attacks.

Notes

Rights