In the rapidly evolving landscape of low-altitude economy, China UAV (Unmanned Aerial Vehicle) technology plays a pivotal role in logistics, urban air mobility, emergency rescue, and public security. The secure and reliable transmission of critical commands from UAV base stations to multiple ground users under stringent latency and covertness constraints has become a pressing challenge. We propose a dual-UAV cooperative covert communication system leveraging Rate-Splitting Multiple Access (RSMA) to address these requirements. One UAV base station (BS) transmits short-packet data to multiple ground users using RSMA, while a companion UAV emits artificial noise to jam the illegal warden (Willie). By analyzing the detection limit of the warden, we derive explicit constraints for covert short-packet communication. We formulate a non-convex optimization problem to maximize the effective covert sum rate, and develop a joint beamforming and rate allocation algorithm based on block coordinate descent and successive convex approximation. Simulation results demonstrate that the proposed cooperative jamming mechanism improves the covert sum rate by approximately four times, and significantly outperforms SDMA and NOMA schemes in terms of both reliability and covertness. This work provides a robust solution for China UAV networks operating in high-sensitivity environments.
The rapid expansion of China UAV applications in low-altitude economy demands ultra-reliable and low-latency communication (URLLC) along with information covertness. Unlike traditional security measures that focus on encryption, covert communication aims to hide the very existence of transmission from adversaries. This is critical for government, military, and sensitive commercial missions where detection of communication activity can lead to catastrophic consequences. However, UAV channels exhibit high mobility, line-of-sight (LoS) dominance, and time-varying fading, making covert communication particularly challenging. Short-packet communication, essential for low-latency control, introduces non-negligible decoding errors and rate penalties due to finite blocklength. Additionally, multi-user interference must be carefully managed. RSMA, as a universal interference management framework, splits each user’s message into a common part and a private part. The common stream is decoded by all users, while private streams are decoded after successive interference cancellation. This flexible structure naturally accommodates covertness and jamming. We integrate RSMA with a cooperative jamming UAV to enhance covert performance. Our work, supported by the National Railway Intelligent Transportation System Engineering Technology Research Center (RITS2025KF01), aims to design a practical solution for China UAV-based URLLC covert systems.
We consider a downlink covert communication system in an urban environment as shown in the system model. The source UAV Alice (equipped with Nt antennas) transmits short-packet data to K single-antenna ground users (Bobs) with the assistance of a cooperative jamming UAV Helper (also Nt antennas). A single-antenna illegal warden Willie attempts to detect whether Alice is transmitting. Due to building obstructions, Alice-Bob links are non-line-of-sight (NLoS) with Rician fading. Helper continuously transmits artificial noise to degrade Willie’s detection performance.
Alice uses RSMA to split each user’s message O_k into common part O_k^c and private part O_k^p. All common parts are combined into a single common stream s_c, and each private part is encoded into private stream s_k (k=1,…,K). The transmitted signal from Alice is: x_Alice = w_c s_c + Σ_{k=1}^K w_k s_k, where w_c and w_k are beamforming vectors for common and private streams, respectively. Helper transmits jamming signal x_Helper = z s_z, where z is the jamming beamforming vector and s_z is unit-power artificial noise.
The received signal at user k is: y_k = h_k^H x_Alice + g_k^H x_Helper + n_k, where h_k ∈ ℂ^{Nt×1} is the channel from Alice, g_k ∈ ℂ^{Nt×1} from Helper, and n_k ~ CN(0,σ²). Users decode using SIC: first decode common stream treating all private streams and jamming as interference, then decode their private stream after removing the common stream. The SINR for common stream at user k is: γ_{c,k} = |h_k^H w_c|² / (Σ_{j=1}^K |h_k^H w_j|² + |g_k^H z|² + σ²). Similarly, private SINR after successful common decoding is: γ_{p,k} = |h_k^H w_k|² / (Σ_{j≠k} |h_k^H w_j|² + |g_k^H z|² + σ²).
Under finite blocklength m and maximum decoding error probability ε, the achievable rate for stream i (i = c or p) at user k is approximated by Polyanskiy et al.: R_{i,k} = log₂(1+γ_{i,k}) – √(V_{i,k}/m) Q^{-1}(ε) / ln2, where V_{i,k} = 1 – 1/(1+γ_{i,k})² is the channel dispersion. The common stream must be decodable by all users, so its total rate must satisfy: Σ_{k=1}^K α_k ≤ min_k R_{c,k}, where α_k is the portion of common rate allocated to user k. The total effective covert rate for user k is: R_k = R_{p,k} + α_k.
| Parameter | Symbol | Value |
|---|---|---|
| UAV altitude | H | 100 m |
| Number of users | K | 3 |
| Antennas at Alice/Helper | Nt | 4 |
| Max transmit power | P_max | 30 dBm |
| Path loss exponent | η | 2.2 |
| Rician factor | κ | 3 dB |
| Reference channel gain | β0 | -30 dB |
| Noise power | σ² | -100 dBm |
| Blocklength | m | 500 |
| Covertness tolerance | ε_c | 0.05 |
Willie’s detection is formulated as a binary hypothesis test. Under H0 (Alice silent), the received signal at Willie is: y_w = g_w^H z s_z + n_w, where g_w is channel from Helper, n_w ~ CN(0,σ²). Under H1 (Alice transmitting), y_w = h_w^H (w_c s_c + Σ w_k s_k) + g_w^H z s_z + n_w. Willie aims to minimize detection error probability ζ = P_FA + P_MD. Using Pinsker’s inequality, we approximate the covertness constraint via Kullback-Leibler (KL) divergence: D(P0||P1) ≤ 2(ε_c)². Both distributions are zero-mean complex Gaussian with variances λ0 and λ1: λ0 = |g_w^H z|² + σ², λ1 = |h_w^H w_c|² + Σ |h_w^H w_k|² + |g_w^H z|² + σ². The KL divergence simplifies to: D(P0||P1) = ln(λ1/λ0) + λ0/λ1 – 1. To guarantee covertness, we require: D(P0||P1) ≤ δ, where δ = 2(ε_c)².
We formulate the optimization problem (P1) to maximize the total effective covert sum rate Σ R_k subject to: (C1) KL divergence constraint, (C2) common stream decodability, (C3) Alice’s power constraint Σ(||w_c||²+Σ||w_k||²) ≤ P_A, (C4) Helper’s power constraint ||z||² ≤ P_H, (C5) α_k ≥ 0. The problem is non-convex due to finite blocklength rate expressions and coupled variables. We propose an alternating iterative algorithm based on block coordinate descent (BCD) and successive convex approximation (SCA).
First, we approximate the non-convex finite-blocklength rate using first-order Taylor expansion. For any variable γ, define f(γ) = log₂(1+γ) – c √(V(γ)) where c = Q^{-1}(ε)/ (ln2 √m). At iteration n, we linearize -c√(V(γ)) around γ^{(n)} to obtain a concave lower bound: f(γ) ≥ f(γ^{(n)}) + f'(γ^{(n)})(γ – γ^{(n)}). The gradient f'(γ) is computed analytically. Second, we handle the non-convex KL divergence constraint. We rewrite D(P0||P1) ≤ δ by introducing auxiliary variables τ and β. Through expansion and SCA, we transform (C1) into three convex constraints:
Constraint (35): ln(τ) + λ0/τ – 1 ≤ δ
Constraint (36): λ1 ≤ τ
Constraint (37): β ≤ λ0, with additional linear approximations.
The resulting subproblem (P2) for given α is a convex optimization over w_c, {w_k}, z, τ, β, solvable via CVX. Then we update α by solving a linear program (P3): maximize Σ α_k subject to α_k ≤ R_{c,k} and Σ α_k ≤ min_k R_{c,k}. We alternate between (P2) and (P3) until convergence. The algorithm is summarized as follows:
| Step | Description |
|---|---|
| 1 | Input: CSI, tolerance δ, max iterations L_max; Initialize W^{(0)}, z^{(0)}, α^{(0)}, n=0. |
| 2 | Repeat: |
| 3 | Compute Taylor coefficients for short-packet rates at current point. |
| 4 | Compute linear approximations for KL constraint. |
| 5 | Solve (P2) for W^{(n+1)}, z^{(n+1)} using CVX. |
| 6 | Update common stream rates: R_{c,k} from (P2) solution; solve (P3) for α^{(n+1)}. |
| 7 | n = n+1; compute objective increment Δ. |
| 8 | Until Δ < δ or n ≥ L_max. |
| 9 | Output: optimal W*, z*, α*. |
We conducted extensive simulations to evaluate the proposed scheme. The simulation setup follows Table 1. Alice at (0,0,100) m, Helper at (100,0,100) m, users randomly distributed within a circle of radius 50 m centered at (0,0,0), Willie at (200,0,0). Path loss exponent 2.2, Rician factor 3 dB, blocklength m=500, covertness threshold ε_c=0.05. We compare RSMA with SDMA (space division multiple access) and NOMA (non-orthogonal multiple access) benchmarks, as well as a non-jamming scenario.
The convergence behavior of the proposed algorithm is shown in Figure 1 (convergence plot). The algorithm converges within 6 iterations for both Nt=4 and Nt=8. The RSMA curve always stays above SDMA and NOMA. Increasing antennas improves the final sum rate but does not affect convergence speed.
Figure 2 illustrates the impact of Alice’s maximum transmit power P_A on covert sum rate. With cooperative jamming, the RSMA scheme achieves about 4 times higher rate than without jamming at P_A=30 dBm. RSMA outperforms SDMA and NOMA across all power levels, especially in the high-power regime where covertness constraints become dominant.
Figure 3 shows the effect of blocklength m. As m increases, the finite-blocklength penalty decreases, and all schemes approach the Shannon limit. RSMA consistently achieves the highest rate, demonstrating robustness to short-packet penalties. In the low-blocklength region (m<300), the rate of RSMA is more than 20% higher than NOMA.
Figure 4 depicts the influence of covertness tolerance ε_c. For very strict covertness (ε_c < 0.02), all schemes achieve near-zero rate. As ε_c increases to 0.05, the rate increases sharply. RSMA shows the steepest improvement, confirming its efficiency in balancing covertness and throughput. At ε_c=0.1, RSMA outperforms SDMA and NOMA by 12.5% and 28.6%.
Figure 5 presents the trade-off between blocklength and decoding error probability ε. Given a target ε, RSMA requires the shortest blocklength. For ε=10^{-4}, RSMA needs m≈400, while NOMA requires m≈600 and SDMA m≈800. This demonstrates the latency advantage of RSMA in China UAV URLLC scenarios.
In summary, we have proposed a dual-UAV cooperative RSMA covert communication system tailored for China UAV networks in high-sensitivity low-altitude economy missions. By jointly optimizing beamforming vectors, jamming design, and RSMA rate allocation, we maximize the effective covert sum rate under finite-blocklength and strict covertness constraints. The proposed BCD-SCA algorithm efficiently solves the non-convex problem. Numerical results demonstrate that introducing cooperative jamming improves the covert sum rate by a factor of four, and RSMA significantly outperforms SDMA and NOMA in terms of both reliability and covertness. This work provides a foundational solution for URLLC covert communications in China UAV applications, with potential extensions to dynamic trajectory optimization and imperfect CSI scenarios in future research.