Effect of Sampling Rate on Transient Based RF Fingerprinting Selçuk Taşcıoğlu 1 , Memduh Köse 2 , and Ziya Telatar 1 1 Department of Electrical and Electronics Engineering Ankara University selcuk.tascioglu@eng.ankara.edu.tr, ziya.telatar@ankara.edu.tr 2 Computer Sciences Research and Application Center Ahi Evran University memduh.kose@ahievran.edu.tr Abstract In this paper, effect of sampling rate on the performance of transmitter identification system using transient-based RF fingerprints is considered. Two different existing RF fingerprinting techniques have been employed to investigate the performance of a transmitter identification system by using experimental data collected at a high sampling rate. Decimation was carried out to analyze the effect of lower sampling rates. It has been shown that transient-based RF fingerprinting methods can be effectively used for identification of wireless transmitters at low sampling rates. 1. Introduction RF fingerprints are defined as the unique characteristics of transmitters caused by their radio circuitry. These unique characteristics can be employed for the identification of wireless devices. RF fingerprinting methods have been employed for identification of several wireless devices, e.g. VHF transmitters [1]-[3], WiFi [4]-[8] and UMTS [9], [10] transceivers. An overview of transmitter identification systems based on RF fingerprinting is presented in [11]. The main stages of an identification system based on RF fingerprinting are defined as signal detection, feature extraction, and classification. After detecting identification signals, such as transients and preambles, from the transmitted signals, distinctive features are extracted from the detected signals and employed to classify transmitters. The identification systems using steady state characteristics such as preambles can exploit the prior information about the known signals. On the other hand, the identification systems using transient characteristics have the advantage that the unintended transients are transmitted before settling down to a steady state condition for all types of wireless devices. In [9], transient-based RF fingerprinting methods are claimed to require extremely high sampling rates to extract features from transient signals without providing experimental or simulation results. This has been regarded as a major disadvantage of the transient based RF fingerprinting methods in [10], [12], [13]. In [12], transient detection stage, as well as feature extraction stage, is claimed to require high sampling rate due to its relatively small duration compared to steady state signal regions. However, none of these works deals with verifying the claim about the high sampling rate requirement of transient based RF fingerprinting methods. In this paper, effect of sampling rate on the classification performance of a transient-based transmitter identification system is investigated. For this purpose, two different existing RF fingerprinting methods, which are based on instantaneous amplitude responses of turn-on transient signals [4], have been employed. It has been shown through experimental data that, contrary to the claim in the literature, high sampling rate is not a requirement to identify wireless transmitters through transient- based RF fingerprinting. The organization of the paper is as follows: In Section 2, a brief description of transmitter identification system using transient based RF fingerprints is presented. In Section 3, decimation process applied to experimental data is explained. Performance evaluation results of the transient based RF fingerprinting method at low sampling rates is presented in Section 4. Our comments on the effect of sampling rate on transient based RF fingerprinting are given in Section 5. Finally, section 6 concludes the paper. 2. Transmitter Identification Through Transient Based RF Fingerprinting The transmitter identification procedure using transient based RF fingerprints is depicted in Fig.1. Sampled baseband signals including transients following channel noise are applied to a detector to find the transient signals. In this study, we used a Bayesian ramp detector [14], to obtain transient signals. This algorithm estimates the transient starting points based on a likelihood function constructed under Gaussian noise assumption. In [15], transient detection performance using this detector was evaluated at different SNR levels and it was demonstrated that the Bayesian ramp detector can be used with a high accuracy for the SNR levels above 10 dB. In the same study, the authors also investigated the effect of detection errors on overall identification system performance for varying SNRs, where SNR levels of transient signals were calculated using the approximation in [16]. In feature extraction stage, we used instantaneous amplitude responses (Amplitude features) [4] and their dimensionally reduced forms obtained by using principal component analysis (PCA features) [4]. At the last stage, a probabilistic neural network (PNN) classifier was used to classify the transmitters by using the extracted features. PNN classifiers have been widely used for classification of transmitters [2], [4], [7], [15]. In [7], the performance of the PNN classifier was compared to a k-nearest neighbor (kNN) classifier in a transmitter classification problem and observed that the PNN classifier outperforms the kNN classifier for varying SNR levels and training sample sizes.