Transfer Learning for Device Fingerprinting with Application to Cognitive Radio Networks

Abstract

Primary user emulation (PUE) attacks are an emerging threat to cognitive radio (CR) networks in which malicious users imitate the primary users (PUs) signals to limit the access of secondary users (SUs). Ascertaining the identity of the devices is a key technical challenge that must be overcome to thwart the threat of PUE attacks. Typically, detection of PUE attacks is done by inspecting the signals coming from all the devices in the system, and then using these signals to form unique fingerprints for each device. Current detection and fingerprinting approaches require certain conditions to hold in order to effectively detect attackers. Such conditions include the need for a sufficient amount of fingerprint data for users or the existence of both the attacker and the victim PU within the same time frame. These conditions are necessary because current methods lack the ability to learn the behavior of both SUs and PUs with time. In this paper, a novel transfer learning (TL) approach is proposed, in which abstract knowledge about PUs and SUs is transferred from past time frames to improve the detection process at future time frames. The proposed approach extracts a high level representation for the environment at every time frame. This high level information is accumulated to form an abstract knowledge database. The CR system then utilizes this database to accurately detect PUE attacks even if an insufficient amount of fingerprint data is available at the current time frame. The dynamic structure of the proposed approach uses the final detection decisions to update the abstract knowledge database for future runs. Simulation results show that the proposed method can improve the performance with an average of 3.5% for only 10% relevant information between the past knowledge and the current environment signals.