Identification of wireless users via power amplifier imperfections

Identification of Wireless Devices of Users Who Actively Fake Their RF Fingerprints With Artificial Data Distortion

“…When such a distortion is employed, the correctly detected data symbols, chosen at the receiver as elements of the constellation, differ from the symbols at the transmitter directly before the pulse-shaping and the exposure to the PA nonlinearities. This intentional distortion, while causing only slight signal quality degradation (see SectionV), causes significant degradation of the identification methods from [1] and [2], which makes the attack very attractive for the strong adversary. In particular, in [1] and [2] we assumed that for the PA, which dominates the transmitter's nonlinearity, the input samples are accessible, as they can be reconstructed from the correctly decoded data at the receiver if the rest of the transmitter chain is assumed linear.…”

“…Exploitation of imperfections of hardware caused by inaccuracies of production processes is especially attractive for identification purposes, because it makes identification independent from the location of wireless users, as opposed to the methods based on channel properties [6]- [9] that require a strong assumption on users' stationarity. Physical-layer identification techniques that exploit hardware imperfections can generally be divided into two groups: transient signal techniques [10]- [17] and steady state signal techniques [1], [2], [18]- [21]. A transient is a radio emission generated while the power of the output of an radio frequency (RF) transmitter goes from zero to the level required for data communication.…”

“…Also, accurate estimation of the start and end times of the transients is challenging. Recently, device identification techniques based on the analysis of slight variations of steady state, modulated signals have been proposed [1], [2], [18]- [21]. Brik et al [18] and later Candore et al [20] used machine learning techniques on collected modulation data to train classifiers that are then able to distinguish wireless cards.…”

“…Brik et al [18] and later Candore et al [20] used machine learning techniques on collected modulation data to train classifiers that are then able to distinguish wireless cards. In [1], [2] we introduced a model-based approach that provides very good identification performance and allows for verifiable accuracy of the identification decisions.…”

“…However, a strong adversary, aware of the fingerprinting methods, could inject slight distortions to the digital data signal before the signal is exposed to the transmitter's nonlinearities, on which the RF fingerprint is based. This slight, artificial distortion while still allowing for reliable data transmission (see Section V), would change the character of the total distortion observable at the receiver over time, and hence the character of the fingerprint, which would significantly degrade performance of the steady state signal based methods [1], [2], [18]. In particular in [22] authors report thwarting the identification methods of Brik et al [18], with success rate close to 100%, with a simple adjustment of the carrier frequency of the masquerading device and with digital shrinking/expanding of the constellation symbols' positions.…”