Deep-Learning-Assisted Optical Transmitter Fingerprint Identification Based on the Constellation Diagram

For enhancing the security of coherent optical communication systems, an optical transmitter identification method based on constellation diagrams is proposed. Addressing the limitations of existing optical transmitter identification methods in adapting to higher order modulation formats, the constellation diagram is employed as the optical transmitter fingerprint. Different optical transmitters can be distinguished by the amplitude features and phase features in the constellation diagram. Four convolutional neural networks (CNNs) are applied for extracting these features to classify optical transmitters, including CX-LeNet, CX-AlexNet, VGG-16, and ResNet-18. Recognition performance of a network is evaluated with the help of recognition accuracy and precision-recall (PR) curves. The employment of the gradient-weighted class activation mapping (grad-CAM) explainer reflects the attention of the network. In the back-to-back system, the accuracy of classification decreases with the increasing symbol rate and increases with the increasing number of symbols. In the 100-km coherent optical communication system with the symbol rate of 25 GBaud, the accuracy of identifying six optical transmitters transmitting 16QAM signals with the VGG-16 network is 100%. Further simulation proves that the proposed method is still feasible in high-speed and long-distance coherent optical communication systems. Effective detection of unauthorized samples proves the ability of the proposed method to enhance the security of coherent optical communication systems.