|
Similar to radio frequency fingerprints, the light emitted by light-emitting diode (LED) contains subtle features of the device itself. These subtle features are only related to the physical layer characteristics of the device itself and are intrinsic features of LED, which are difficult to forge and can be used for LED identification. LED fingerprint identification is of great significance in the fields of visible light indoor positioning and secure access to visible light 5G communication networks. This paper experimentally demonstrated LED light fingerprint identification on three commercial LED lamps of the same model and the same batch. To simulate practical application scenarios, this paper conducted 25 spectral measurements on each LED lamp under indoor ambient light noise conditions. Before each spectral measurement, the relative distance and angle between the LED and the spectrometer were randomly adjusted. In response to the one-dimensional data structure of the measured spectrum, this paper replaces the original two-dimensional convolutional kernel with a one-dimensional convolutional kernel based on the classic VGG19 network, and adds an adaptive average pooling layer in front of the fully connected layer, proposing an improved VGG19-1D identification algorithm. The improved algorithm was trained from scratch in the training dataset, and after 1200 epochs of training, the test identification accuracy of commercial LED lamps reached 91.30%. The experimental results demonstrated that reliable identification can be achieved based on emitted light spectrum for commercial LED lamps of the same model and the same batch. The improved algorithm has significantly higher identification accuracy than traditional machine learning methods such as support vector machine and random forests.
|
