Secure and calibrated federated QoT estimation over heterogeneous optical-network domains

Quality-of-transmission (QoT) estimation supports lightpath admission control and reconfiguration in optical transport networks, but collaborative model training is hindered when QoT records are distributed across operators, regions, or other administrative domains that cannot pool raw data. We study Domain-Adaptive Secure Federated QoT (DA-SFQ), a federated QoT workflow for heterogeneous optical-network domains that share the same feature definitions but differ in topology, load, fiber/space-division multiplexing (SDM) configuration, switching paradigm, and operating regime. DA-SFQ shares a representation backbone while keeping client-private classification and regression heads local for QoT feasibility screening and optical signal-to-noise ratio (OSNR) estimation. Secure aggregation (SA) hides individual backbone updates from the coordinating server, and client-side calibration with recall-constrained threshold selection reports the false-acceptance risk of QoT-failing lightpaths. Evaluations on a public SDM QoT benchmark with 23,616 samples and K ∈{2,3,6,16} collaboration settings, complemented by a K =2 cross-site prototype, show that DA-SFQ + SA achieves worst-domain feasibility accuracy of 0.988/0.975 for K =2/ K =3 and macro accuracy of 0.897/0.880 for the more heterogeneous K =6/ K =16 settings. For OSNR regression, DA-SFQ + SA obtains in-domain macro mean absolute error (MAE) of 1.52, 0.14, and 1.57 dB for K =2,3,6, respectively. SA leaves the QoT metrics close to the plain mode; under the emulation-based payload accounting used here, the reported per-round upload approximately doubles. New-domain stress tests further show that high score separability does not by itself guarantee stable admission thresholds under severe domain mismatch, motivating domain-aware calibration before deployment.