End-to-end deep learning framework for key-free physical-layer security in WDM-RoF

The growing demand for high-throughput and secure fronthaul in future networks calls for advanced physical-layer security (PLS) for wavelength-division multiplexing radio-over-fiber (WDM-RoF) systems. This paper introduces an end-to-end (E2E) deep learning (DL) framework that achieves PLS for WDM-RoF links without relying on explicit cryptographic keys. Security is enforced by two concurrent mechanisms: (1) dynamic interlocking of jointly trained transmitter (TransNN) and receiver (ReceivNN) neural network pairs, and (2) the exclusive binding of each TransNN-ReceivNN pair to its wavelength channel's physical characteristics. Reliable demodulation is obtained only when both conditions are satisfied. Numerical simulations on a four-wavelength WDM-RoF system over a 20-km fiber show that authorized connections maintain a bit error rate (BER) below the 7% forward error correction (FEC) threshold. Conversely, mismatched model pairings result in BER values of around 0.3, whereas deploying a correct model pair on non-designated wavelength channels results in BERs approaching 0.5. The proposed E2E approach embeds security directly into the physical layer, eliminating the need for key management, synchronization, and auxiliary encryption hardware. This approach thereby offers a lightweight yet robust PLS solution for WDM-RoF systems.