Artículo de revista
Convolution based feature extraction for edge computing access authentication
Fecha
2020Registro en:
IEEE Transactions on Network Science and Engineering, Vol. 7, No. 4, October-December 2020
10.1109/TNSE.2019.2957323
Autor
Xie, Feiyi
Wen, Hong
Wu, Jinsong
Chen, Songlin
Hou, Wenjing
Jiang, Yixin
Institución
Resumen
In this article, a convolutional neural network (CNN)
enhanced radio frequency fingerprinting (RFF) authentication
scheme is presented for Internet of things (IoT). RFF is a
non-cryptographic authentication technology, identifies devices
through the waveforms of the RF transient signals by processing
received RF signals on the edge server, which places no cost
burden to low-end (low-cost) devices without implementing any
encryption algorithmand meet the demands of the real-time access
authentication in Internet of things. In the new scheme, the
feasibility of extracting features based on one-dimensional
(1D) signal convolution is discussed, referring to the method
of extracting features from CNN, and combining with the
characteristics of signal convolution. A convolution kernel for 1D
signals is designed to extract the feature of signals in order to
reduce training time and ensure classification accuracy. Therefore,
it can improve the accuracy compared with these traditional
algorithms, while saving the training time of updating parameters
repeatedly as the neural network. The accuracy and training time
of thealgorithm are verified in a real signal acquisition system. The
results prove that the novel algorithm can effectively improve the
classification accuracy in low signal-to-noise ratio (SNR), while
keeps the training time in an acceptable range.