This work proposes a secure adaptive federated authentication framework, SAFA‑MZ, for non‑terrestrial networks (NTN). The threat model assumes passive, location‑aware, static eavesdroppers that cannot obtain legitimate channel state information (CSI) but can listen to uplink and downlink signals. To counter such eavesdropping, group‑level authentication tags are embedded into a multi‑layer rate‑splitting multiple access (RSMA) transmission. Private and common signals are jointly beamformed, artificial noise (AN) is injected to degrade the eavesdropper’s observation, and group differential privacy (GDP) protects tag information from inference attacks. Users are grouped by semantic priority, and secrecy spectral efficiency (SSE) is allocated according to information importance.
The design variables include high‑altitude platform station (HAPS) placement, user association, and RSMA power allocation. We formulate an SSE maximization problem under authentication reliability, power limits, coverage, and probabilistic secrecy constraints:
$$ \max_{\mathbf{p},\mathbf{A},\mathbf{x}} \; \mathrm{SSE}(\mathbf{p},\mathbf{A},\mathbf{x}) $$
where $\mathbf{p}$ denotes the power vector, $\mathbf{A}$ the HAPS deployment matrix, and $\mathbf{x}$ the user‑association indicator. Constraints enforce $\Pr\{\text{authentication error}\}\le \epsilon_{\text{auth}}$, $\Pr\{\text{information leakage}\}\le \epsilon_{\text{sec}}$, and per‑HAPS power budget $P_{\max}$. To solve this mixed‑integer non‑convex problem, two algorithms are introduced. The repair‑based cross‑entropy method (RCEM) iteratively samples solutions, repairs infeasible ones, and converges with $O(N^2)$ complexity ( $N$ is the number of users). The graph‑aware advantage actor‑critic (GA2C) maps user association to a graph, extracts neighborhood features via graph convolutional networks, and updates policies with an advantage actor‑critic scheme, achieving $O(N)$ complexity. Simulations under colluding and non‑colluding eavesdroppers show that SAFA‑MZ improves average SSE by up to 135% over single‑connect transmission and by 21% compared to a scheme without AN. These results confirm the scalability and security of the proposed solution in dynamic NTN scenarios.
Review: SAFA‑MZ tightly integrates authentication tags with RSMA and leverages graph reinforcement learning for low‑latency inference, balancing privacy and resource efficiency, and offering a practical pathway for secure NTN access.