Self-adaptive hybrid data-model optimization for secure end-to-end radio-over-fiber transmission.

Abstract

A novel self-adaptive secure end-to-end (E2E) transmission approach is proposed for a radio-over-fiber (RoF) system. The system integrates deep learning (DL) and traditional models across the transmitter, channel, and receiver, forming an E2E transmission framework. The encryption function of the system is embedded into modulation (TransNN) and demodulation (ReceivNN) via E2E optimization. Training-phase randomization and noise perturbations ensure incompatibility between modulation and demodulation models across different training rounds. An adversarial training strategy enhances physical-layer security by adapting the demodulation model to the legal channel while restricting its effectiveness on illegal ones. Numerical simulations indicate that under white-box attacks, only the matched ReceivNN correctly demodulates TransNN signals, while under gray-box attacks, ReceivNN demodulation performance degrades due to mismatched channel conditions. These results validate the scheme's effectiveness against both white-box and gray-box attacks, offering a secure and adaptive solution for RoF systems.