Resistance of a Pump-Keyed Multi-Core Fiber Channel to Learning-Based Known-Plaintext Attacks: Identifiability Limits and Dimensional Scaling

Physical layer protection can complement conventional cryptography, but its contribution under repeated key use depends on whether an adversary can learn the channel’s input–output map. We evaluate a pump-keyed multi-core fiber (MCF) architecture in a physically parameterized simulation under full observation (FO) and a partial, attenuated, noisier wiretap-like observation (WT). A fixed-key linear baseline is rapidly learnable under FO. A key-controlled nonlinear phase construction (P2) increases error for the evaluated estimators at small known-plaintext budgets. Extending the evaluation to 50,000 known-plaintext vectors changes that conclusion: at 37 cores and the default nonlinear coefficient gamma_NL = 2, a multilayer perceptron reaches mean BER 0.033, crossing the BER < 0.05 recovery criterion on all three seeds. Increasing the core count to 61 or doubling gamma_NL prevents a crossing within the tested data and training budgets. Doubling gamma_NL raises mean Bob BER from 0.000014 to 0.000471, with a maximum of 0.001007 across the three keys. P3 also remains above the recovery criterion at 50,000 pairs at a larger, key-dependent receiver cost. The core count and nonlinear strength sweeps establish an empirical trend under the specified protocol; they do not determine a general scaling law or a lower bound on the data required by stronger attackers. The results distinguish linear identification from nonlinear decoder learning and do not establish information theoretic secrecy or formal cryptographic security.