Wireless Signal Fngerprint Extraction Based on Convolutional Deep Belief Network

Signal Fingerprint is a physical layer based wireless security algorithm desinged for recognizing legal devices. The non-ideality of transmitting circuit will lead to small but unique distortions in the signal emitted by this circuit. Signal fingerprint is aiming at extracting those unique distortions from received signal in order to identify the related radiofrequency circuit and wireless device. Traditional signal fingerprint extraction depends on carefully designed formula whose application scope is narrow and the accuracy cannot be guaranteed. Supervised deep learning based signal fingerprint extraction enjoys high accuracy and can be deployed in many different circumstances. But the large amount labeled data required in training stage is a strong requirement that cannot be easily met. Besides, the network might be sensitive to noise. To solve those problems, this paper proposes a Convolutional Deep Belief Network (CDBN) based signal fingerprint extraction algorithm. The signal fingerprint extraction is performed by network itself with the spectrum of signal as input. So the accuracy can be improved and application scope can be widen. Due to the unsupervised essence of CDBN, the constraint of labeled data is relieved.