Security threats to signal classifiers using self-organizing maps

Spectrum sensing is required for many cognitive radio applications, including spectral awareness, interoperability, and dynamic spectrum access. Previous work has demonstrated the ill effects of primary user emulation attacks, and pointed out specific vulnerabilities in spectrum sensing that uses feature-based classifiers. This paper looks specifically at the use of unsupervised learning in signal classifiers, and attacks against self-organizing maps. By temporarily manipulating their signals, attackers can cause other secondary users to permanently mis-classify them as primary users, giving them complete access to the spectrum. In the paper we develop the theory behind manipulating the decision regions in a neural network using self-organizing maps. We then demonstrate through simulation the ability for an attacker to formulate the necessary input signals to execute the attack. Lastly we provide recommendations to mitigate the efficacy of this type of attack.