Noise-Dependent Eavesdropping Strategies in BB84: A Quantum Circuit Learning Approach

Quantum circuit learning (QCL) has recently been proposed as a framework for optimizing eavesdropping attacks on the BB84 protocol. In this work, we investigate how learned attacks behave under realistic noisy conditions by systematically analyzing multiple single-qubit noise models, including depolarizing, bit-flip, phase-flip, amplitude damping, phase damping, and combined error processes. We first validate the variational framework in the noiseless regime against the analytical phase-covariant cloning machine (PCCM) frontier, and then extend the analysis to noisy channels, reproducing known results and generalizing them across distinct noise families. Our results show that learned attacks do not respond uniformly to noise and that deviations from PCCM-like behavior depend strongly on the structure of the underlying noise rather than solely on its magnitude. In particular, noise processes that introduce basis-dependent effects lead to asymmetric attack strategies, while more symmetric channels tend to preserve the structure of optimal symmetric attacks. These findings indicate that commonly used aggregate metrics, such as quantum bit error rate (QBER), may be insufficient to fully characterize security in practical QKD deployments. The results provide insight into how physical-layer impairments in real communication systems influence the information-disturbance tradeoff, contributing to amore realistic assessment of quantum communication security.