Robust GSNR estimation in dynamic C + L-band optical networks: benchmarking a DNN and an Edge-GAT under adversarial attacks

Abstract

In dynamic $\text{C}+\text{L}$-band optical networks, digital twins (DTs) enable accurate quality of transmission (QoT) estimation, serving as a cornerstone for autonomous network management. However, reliable estimation becomes increasingly challenging in large-scale topologies, where conventional deep neural networks (DNNs) ignore optical multiplex section (OMS)-level topological dependencies and show limited robustness under input-power perturbations. To address this issue, we propose a robust generalized signal-to-noise ratio (GSNR) estimation framework based on an edge-based graph attention network (Edge-GAT) with joint adversarial training, using GNPy-generated OMS-level full-network snapshots. The proposed framework is evaluated on both six-node and NSFNet topologies. Results show that the robust Edge-GAT achieves clean/adversarial mean absolute errors (MAEs) of 0.028/0.035 dB on the six-node topology and 0.036/0.051 dB on NSFNet. These results demonstrate the effectiveness of topology-aware robust training for reliable DT-assisted GSNR estimation in dynamic $\text{C}+\text{L}$ optical networks.