The relentless evolution towards the Sixth-Generation (6G) mobile communication network imposes unprecedented demands on spectral efficiency, energy efficiency, ultra-low latency, and ubiquitous coverage. In this context, Unmanned Aerial Vehicle (UAV) or drone communication has emerged as a cornerstone technology, primarily due to its ability to establish high-quality Line-of-Sight (LoS) links. This capability is crucial for ensuring high data-rate transmissions, making UAVs indispensable for 6G and future network paradigms. China’s rapid advancements in UAV drone technology position it at the forefront of integrating aerial platforms into next-generation networks. UAVs, functioning as aerial base stations, offer unparalleled advantages such as controllable mobility, rapid deployment, and operational flexibility. These attributes significantly enhance network performance by improving coverage, especially in challenging scenarios like emergency communications, remote area service, and maritime operations.
However, the open and broadcast nature of wireless channels in UAV communications renders them highly vulnerable to eavesdropping attacks, posing a critical challenge to physical layer security (PLS). Furthermore, serving multiple ground users simultaneously introduces severe inter-user interference, which can drastically degrade system performance. To address the interference management challenge, advanced multiple access schemes beyond Orthogonal Multiple Access (OMA) are essential. Rate-Splitting Multiple Access (RSMA) has recently gained significant attention as a powerful and flexible framework. In RSMA, each user’s message is split into a common part, decodable by all users, and a private part, intended for a specific user. This strategy creates an elegant balance between fully decoding interference and treating it as noise, thereby unifying and generalizing conventional schemes like OMA, Space-Division Multiple Access (SDMA), and Non-Orthogonal Multiple Access (NOMA). Integrating RSMA with UAV drone networks is therefore a promising direction for future secure communications.
Concurrently, the advent of Intelligent Reflecting Surfaces (IRS) offers a revolutionary approach to proactively tailor the wireless propagation environment. An IRS is a planar array composed of numerous low-cost, passive reflecting elements, each capable of independently adjusting the phase shift (and potentially amplitude) of the incident signal. By smartly configuring these phase shifts, the IRS can constructively combine signal paths toward legitimate users while destructively canceling them toward potential eavesdroppers, thereby enhancing secrecy performance without requiring additional energy-intensive radio frequency chains. The synergy between IRS and UAV drone systems is particularly potent: the UAV’s mobility can be exploited to achieve favorable LoS conditions with the IRS, which in turn can extend coverage and enhance security for ground users, especially in complex urban or obstructed environments prevalent in many regions, including China.
While existing research has explored IRS-assisted secure communications and UAV trajectory optimization separately, the integration of IRS, UAV drones, and RSMA for secrecy maximization against multiple eavesdroppers remains a nascent area. Most prior works focus on minimizing transmit power or maximizing the minimum secrecy rate for a single eavesdropper, often using NOMA. This paper addresses this gap by investigating the joint optimization of transmit precoding, common rate allocation, IRS phase shifts, and UAV positioning to maximize the sum secrecy rate of an IRS-assisted UAV-RSMA system in the presence of multiple eavesdroppers. The main contributions are summarized as follows.
System Model and Problem Formulation
We consider a downlink communication system where a UAV base station (BS) equipped with $M$ antennas serves $K$ single-antenna legitimate users. A fixed IRS with $N$ passive reflecting elements is deployed on a building facade to assist the communication. Furthermore, there exist $J$ single-antenna eavesdroppers attempting to intercept the confidential messages. The sets of legitimate users and eavesdroppers are denoted by $\mathcal{K} = \{1, …, K\}$ and $\mathcal{J} = \{1, …, J\}$, respectively. The UAV’s horizontal position is denoted by $\mathbf{q} = (x, y)$ at a fixed altitude $H$. The IRS’s location is $\mathbf{r} = (x_r, y_r, z_r)$. The locations of user $k$ and eavesdropper $j$ are $\mathbf{w}_k = (x_k, y_k, 0)$ and $\mathbf{w}_j = (x_j, y_j, 0)$, respectively.
The channel responses from the UAV to user $k$, the UAV to eavesdropper $j$, the UAV to the IRS, the IRS to user $k$, and the IRS to eavesdropper $j$ are denoted by $\mathbf{h}_{u,k}^H \in \mathbb{C}^{1 \times M}$, $\mathbf{h}_{u,j}^H \in \mathbb{C}^{1 \times M}$, $\mathbf{G} \in \mathbb{C}^{N \times M}$, $\mathbf{h}_{r,k}^H \in \mathbb{C}^{1 \times N}$, and $\mathbf{h}_{r,j}^H \in \mathbb{C}^{1 \times N}$, respectively. We assume Rician fading channels to capture both LoS and non-LoS components, a model particularly relevant for UAV drone channels in China’s diverse terrain. The overall channel from the UAV to a receiver (user or eavesdropper) via the IRS is the concatenation of the UAV-IRS link and the IRS-receiver link, controlled by the IRS reflection matrix. The IRS reflection matrix is $\boldsymbol{\Theta} = \text{diag}(\mathbf{u})$, where $\mathbf{u} = [e^{j\theta_1}, …, e^{j\theta_N}]^T$ and $\theta_n \in [0, 2\pi)$ is the phase shift of the $n$-th element.
The UAV employs RSMA to serve the $K$ users. The message $W_k$ for user $k$ is split into a common part $W_c^k$ and a private part $W_p^k$. All common parts $\{W_c^1, …, W_c^K\}$ are combined into a single common message $W_c$, which is encoded into a common stream $s_c$. Each private part $W_p^k$ is encoded into a private stream $s_k$. Thus, the vector of streams to be transmitted is $\mathbf{s} = [s_c, s_1, …, s_K]^T \in \mathbb{C}^{(K+1)\times1}$, with $\mathbb{E}\{\mathbf{s}\mathbf{s}^H\}=\mathbf{I}$. The streams are linearly precoded by a precoding matrix $\mathbf{P} = [\mathbf{p}_c, \mathbf{p}_1, …, \mathbf{p}_K] \in \mathbb{C}^{M \times (K+1)}$, where $\mathbf{p}_c$ and $\mathbf{p}_k$ are the precoding vectors for the common and $k$-th private streams, respectively. The transmitted signal is $\mathbf{x} = \mathbf{P}\mathbf{s}$.
The received signals at legitimate user $k$ and eavesdropper $j$ are given by:
$$y_k = \mathbf{h}_k^H \mathbf{x} + n_k, \quad y_j = \mathbf{h}_j^H \mathbf{x} + n_j,$$
where $\mathbf{h}_k^H = \mathbf{h}_{u,k}^H + \mathbf{h}_{r,k}^H \boldsymbol{\Theta} \mathbf{G}$ and $\mathbf{h}_j^H = \mathbf{h}_{u,j}^H + \mathbf{h}_{r,j}^H \boldsymbol{\Theta} \mathbf{G}$ are the aggregated channels, and $n_k, n_j \sim \mathcal{CN}(0, \sigma^2)$ are additive white Gaussian noise.
At user $k$, the common stream $s_c$ is decoded first by treating all private streams as noise. The achievable rate for the common stream at user $k$ is $R_{c,k} = \log_2(1 + \Gamma_{c,k})$, with
$$\Gamma_{c,k} = \frac{|\mathbf{h}_k^H \mathbf{p}_c|^2}{\sum_{i \in \mathcal{K}} |\mathbf{h}_k^H \mathbf{p}_i|^2 + \sigma^2}.$$
After successfully decoding and removing $s_c$, user $k$ decodes its private stream $s_k$. The achievable rate for the private stream at user $k$ is $R_{k} = \log_2(1 + \Gamma_{k})$, with
$$\Gamma_{k} = \frac{|\mathbf{h}_k^H \mathbf{p}_k|^2}{\sum_{i \in \mathcal{K}, i \neq k} |\mathbf{h}_k^H \mathbf{p}_i|^2 + \sigma^2}.$$
The common rate $R_c$ must satisfy $R_c \leq \min_{k \in \mathcal{K}} R_{c,k}$ to ensure successful decoding by all users. $R_c$ is then shared among users, with $C_k$ denoting the portion allocated to user $k$ such that $\sum_{k \in \mathcal{K}} C_k = R_c$.
For secrecy, we consider the worst-case scenario where the eavesdroppers can collude. The eavesdropping rates for the common and private streams are:
$$R_{c,j}^{e} = \log_2(1 + \gamma_{c,j}), \quad R_{k,j}^{e} = \log_2(1 + \gamma_{k,j}),$$
where
$$\gamma_{c,j} = \frac{|\mathbf{h}_j^H \mathbf{p}_c|^2}{\sum_{i \in \mathcal{K}} |\mathbf{h}_j^H \mathbf{p}_i|^2 + \sigma^2}, \quad \gamma_{k,j} = \frac{|\mathbf{h}_j^H \mathbf{p}_k|^2}{\sum_{i \in \mathcal{K}, i \neq k} |\mathbf{h}_j^H \mathbf{p}_i|^2 + |\mathbf{h}_j^H \mathbf{p}_c|^2 + \sigma^2}.$$
The achievable secrecy rate for user $k$ is then:
$$R_k^{sec} = \left[ C_k – \max_{j \in \mathcal{J}} R_{c,j}^{e} \right]^+ + \left[ R_{k} – \max_{j \in \mathcal{J}} R_{k,j}^{e} \right]^+,$$
where $[x]^+ \triangleq \max(0, x)$. The system’s sum secrecy rate is $R_{sec}^{tot} = \sum_{k=1}^{K} R_k^{sec}$.
Our objective is to maximize $R_{sec}^{tot}$ by jointly optimizing the UAV’s position $\mathbf{q}$, the precoding matrix $\mathbf{P}$, the common rate allocation $\{C_k\}$, and the IRS phase shift vector $\mathbf{u}$. The problem is formulated as:
$$
\begin{aligned}
(\text{P1}): \max_{\mathbf{P}, \{C_k\}, \mathbf{q}, \mathbf{u}} \quad & \sum_{k=1}^{K} R_k^{sec} \\
\text{s.t.} \quad & \sum_{k \in \mathcal{K}} C_k \leq \min_{k \in \mathcal{K}} R_{c,k} \\
& C_k \geq 0, \quad \forall k \in \mathcal{K} \\
& \text{tr}(\mathbf{P}\mathbf{P}^H) \leq P_{max} \\
& \theta_n \in [0, 2\pi), \quad n=1,…,N \\
& x_{min} \leq x \leq x_{max}, \quad y_{min} \leq y \leq y_{max}
\end{aligned}
$$
where $P_{max}$ is the maximum transmit power at the UAV drone. Problem (P1) is highly non-convex due to the coupling of optimization variables in the objective function and constraints, particularly through the complex channel expressions $\mathbf{h}_k$ and $\mathbf{h}_j$, and the non-convex nature of the secrecy rate expressions and IRS phase shift constraints.
Proposed Hierarchical Optimization Solution
To tackle the intractable non-convex problem (P1), we propose a hierarchical optimization framework that decomposes it into an inner block and an outer block. The inner block optimizes the communication resources ($\mathbf{P}, \{C_k\}, \mathbf{u}$) for a given UAV position $\mathbf{q}$, while the outer block optimizes the UAV position $\mathbf{q}$ based on the solution from the inner block.
Inner Block Optimization: Alternating Optimization for Communication Resources
Given $\mathbf{q}$, we optimize $\mathbf{P}, \{C_k\},$ and $\mathbf{u}$ alternately.
1. Optimizing Precoding Vectors and Common Rate Allocation
With fixed $\mathbf{u}$, the subproblem for $\mathbf{P}$ and $\{C_k\}$ is still non-convex. We employ techniques from convex optimization theory to find a stationary point.
Step 1: Reformulation. We introduce slack variables to handle the non-smooth $[ \cdot ]^+$ operation and the complex fractional terms in the rates. Let $\bar{R}_c^{e} = \max_{j} R_{c,j}^{e}$ and $\bar{R}_{k}^{e} = \max_{j} R_{k,j}^{e}$. The secrecy rate for user $k$ can be lower-bounded by focusing on the case where the difference is positive: $R_k^{sec} \geq \tilde{R}_k^{sec} = C_k – \bar{R}_c^{e} + R_{k} – \bar{R}_{k}^{e}$, provided $C_k \geq \bar{R}_c^{e}$ and $R_{k} \geq \bar{R}_{k}^{e}$. We then maximize the lower bound $\sum_k \tilde{R}_k^{sec}$.
Step 2: Handling Non-Convex Rate Constraints. The constraints involving rates like $R_{c,k}$, $R_{k}$, $R_{c,j}^{e}$, and $R_{k,j}^{e}$ are non-convex due to the ratio of quadratic terms. We use the Successive Convex Approximation (SCA) method. For a constraint like $R = \log_2(1 + \frac{|\mathbf{h}^H\mathbf{p}|^2}{I + \sigma^2}) \geq \bar{R}$, where $I$ is interference power, it is equivalent to $\frac{|\mathbf{h}^H\mathbf{p}|^2}{I + \sigma^2} \geq 2^{\bar{R}}-1$. This can be rewritten as a difference of convex (DC) function: $|\mathbf{h}^H\mathbf{p}|^2 – (2^{\bar{R}}-1)(I + \sigma^2) \geq 0$. At a given local point $(\mathbf{p}^{(t)}, \bar{R}^{(t)})$, we can approximate the concave part (e.g., $- (2^{\bar{R}}-1)I$) using its first-order Taylor expansion to obtain a convex second-order cone (SOC) constraint.
Step 3: Final Convex Approximation. By applying SCA to all relevant non-convex constraints related to achievable rates and eavesdropping rates, and introducing necessary auxiliary variables, the subproblem is approximated as a convex problem, specifically a Second-Order Cone Program (SOCP) or a convex Quadratically Constrained Quadratic Program (QCQP). This convex subproblem can be efficiently solved using standard convex optimization solvers like CVX.
2. Optimizing IRS Phase Shifts
With fixed $\mathbf{P}$ and $\{C_k\}$, the subproblem for optimizing $\mathbf{u}$ (or equivalently, $\boldsymbol{\Theta}$) is also non-convex. A standard approach is to use the Semi-Definite Relaxation (SDR) technique.
Step 1: Homogeneous Quadratics Reformulation. Define $\mathbf{v} = [\mathbf{u}^T, 1]^T \in \mathbb{C}^{(N+1)\times 1}$ and $\mathbf{V} = \mathbf{v}\mathbf{v}^H$, which satisfies $\mathbf{V} \succeq 0$ and $\text{rank}(\mathbf{V})=1$. The effective channel gain can be expressed as a linear function of $\mathbf{V}$. For example,
$$|\mathbf{h}_k^H \mathbf{p}_c|^2 = |(\mathbf{h}_{u,k}^H + \mathbf{h}_{r,k}^H \boldsymbol{\Theta} \mathbf{G})\mathbf{p}_c|^2 = \text{Tr}(\mathbf{H}_k \mathbf{p}_c\mathbf{p}_c^H \mathbf{H}_k^H \mathbf{V}),$$
where $\mathbf{H}_k = [\text{diag}(\mathbf{h}_{r,k}^H)\mathbf{G}; \mathbf{h}_{u,k}^H]$. Similar transformations apply to all other channel gain terms $|\mathbf{h}_k^H \mathbf{p}_i|^2$ and $|\mathbf{h}_j^H \mathbf{p}_i|^2$.
Step 2: SDR and Problem Transformation. By substituting all channel gain terms with linear expressions in $\mathbf{V}$, the objective function and constraints become linear or quadratic in $\mathbf{V}$, except for the non-convex rank-one constraint $\text{rank}(\mathbf{V})=1$. Applying SDR, we relax this rank-one constraint, transforming the problem into a convex Semi-Definite Program (SDP) with a linear objective in $\mathbf{V}$, subject to linear matrix inequality (LMI) constraints and the constraint $[\mathbf{V}]_{n,n} = 1$ for $n=1,…,N+1$ (to ensure unit modulus for IRS elements).
Step 3: Solution Recovery. The relaxed SDP is solved to obtain an optimal $\mathbf{V}^*$. If $\mathbf{V}^*$ has rank one, its principal eigenvector provides the optimal $\mathbf{v}^*$ and hence $\mathbf{u}^*$. If not, Gaussian randomization is employed to generate a set of candidate vectors from $\mathbf{V}^*$ and select the one yielding the highest objective value.
The inner block optimization alternates between solving the precoding/common-rate SOCP and the IRS phase-shift SDP until convergence. The overall algorithm for the inner block is summarized in Table 1.
| Step | Action |
|---|---|
| 1 | Initialize $\mathbf{P}^{(0)}, \{C_k^{(0)}\}, \mathbf{u}^{(0)}$. Set iteration index $t=0$. |
| 2 | Repeat |
| 2.1 | Given $\mathbf{u}^{(t)}$, solve the convex approximation (SOCP) for $\mathbf{P}$ and $\{C_k\}$ to obtain $\mathbf{P}^{(t+1)}, \{C_k^{(t+1)}\}$. |
| 2.2 | Given $\mathbf{P}^{(t+1)}, \{C_k^{(t+1)}\}$, solve the SDP for $\mathbf{V}$ (relaxed IRS problem). Recover $\mathbf{u}^{(t+1)}$ via eigenvalue decomposition or Gaussian randomization. |
| 2.3 | Update $t = t+1$. |
| 3 | Until the increase in the sum secrecy rate $R_{sec}^{tot}$ is below a predefined threshold $\epsilon$. |
Outer Block Optimization: UAV Position via Particle Swarm Optimization
The optimization of the UAV drone’s position $\mathbf{q}$ is a non-convex problem with respect to the sum secrecy rate objective, which is a complex, non-differentiable function obtained from the inner optimization. Gradient-based methods are challenging to apply. Therefore, we employ the population-based heuristic Particle Swarm Optimization (PSO) algorithm, which is effective for such black-box optimization problems.
In PSO, a swarm of particles (candidate positions) explores the search space defined by $[x_{min}, x_{max}] \times [y_{min}, y_{max}]$. Each particle $i$ has a position $\mathbf{q}_i$ and a velocity $\mathbf{v}_i$. The particles move according to simple rules that incorporate their own best-known position ($\mathbf{pbest}_i$) and the swarm’s global best-known position ($\mathbf{gbest}$). The update equations for particle $i$ at iteration $\tau$ are:
$$
\begin{aligned}
\mathbf{v}_i^{(\tau+1)} &= \omega \mathbf{v}_i^{(\tau)} + c_1 r_1 (\mathbf{pbest}_i – \mathbf{q}_i^{(\tau)}) + c_2 r_2 (\mathbf{gbest} – \mathbf{q}_i^{(\tau)}) \\
\mathbf{q}_i^{(\tau+1)} &= \mathbf{q}_i^{(\tau)} + \mathbf{v}_i^{(\tau+1)}
\end{aligned}
$$
where $\omega$ is the inertia weight, $c_1$ and $c_2$ are acceleration coefficients, and $r_1, r_2 \sim U(0,1)$ are random numbers. For a given particle position $\mathbf{q}_i$, we run the inner block optimization algorithm (Table 1) to compute the corresponding optimal sum secrecy rate $R_{sec}^{tot}(\mathbf{q}_i)$, which serves as the fitness value. The PSO process continues until convergence or a maximum number of iterations is reached. The final $\mathbf{gbest}$ provides the optimized UAV drone position.
The complete hierarchical algorithm is illustrated in the flowchart below, showing the interaction between the outer PSO loop and the inner alternating optimization loop.
Performance Evaluation and Numerical Results
We conduct extensive simulations to evaluate the performance of the proposed algorithm. The simulation parameters, reflective of scenarios where China UAV drone deployments are pertinent, are listed in Table 2.
| Parameter | Value |
|---|---|
| UAV Altitude ($H$) | 100 m |
| UAV Max Transmit Power ($P_{max}$) | 30 dBm |
| Number of UAV Antennas ($M$) | 4, 6 (varied) |
| Number of Legitimate Users ($K$) | 4 |
| Number of Eavesdroppers ($J$) | 3 |
| Number of IRS Elements ($N$) | 32 |
| Carrier Frequency / Wavelength ($\lambda$) | 2.4 GHz |
| Path Loss at 1 m ($\rho_0$) | -30 dB |
| Path Loss Exponent (UAV-User, UAV-Eaves.) | 3.6 |
| Rician Factor ($\beta$) | 10 dB |
| Noise Power ($\sigma^2$) | -90 dBm |
| PSO Parameters ($\omega, c_1, c_2$, swarm size) | 0.7, 2.0, 2.0, 20 |
We compare the proposed “IRS-RSMA” scheme with the following benchmarks:
- RSMA without IRS: UAV uses RSMA but without IRS assistance ($\boldsymbol{\Theta}=\mathbf{0}$).
- NOMA with IRS: UAV uses conventional NOMA (a special case of RSMA with no common stream) assisted by the IRS.
- NOMA without IRS: Baseline with NOMA and no IRS.
Convergence Behavior
The inner alternating optimization algorithm demonstrates fast convergence, typically within 10-15 iterations, as the SCA and SDR steps successively improve the lower bound of the objective. The outer PSO algorithm converges within 30-40 iterations, showing the effectiveness of the hierarchical approach in finding a good UAV drone position. The overall computational complexity is manageable, dominated by solving the inner SDP, which scales polynomially with the number of IRS elements $N$.
Impact of Transmit Power and UAV Antennas
Figure X (conceptual) shows the sum secrecy rate versus the UAV’s maximum transmit power $P_{max}$. As $P_{max}$ increases, all schemes see an improvement because higher power strengthens the signal at legitimate users. The proposed IRS-RSMA scheme consistently outperforms all benchmarks. The performance gain over “RSMA without IRS” highlights the value of the IRS in shaping the propagation environment to enhance legitimate channels and suppress eavesdropper channels. The significant gain over both NOMA-based schemes underscores the superiority of RSMA’s interference management in this multi-eavesdropper secure communication scenario enabled by the China UAV drone platform.
Similarly, increasing the number of UAV antennas $M$ provides more degrees of freedom for beamforming. Our proposed scheme benefits the most from additional antennas, as it can leverage them for more precise precoding of both common and private streams, while the IRS provides an additional dimension for beamforming in the reflection domain.
Impact of IRS Capability
Figure Y (conceptual) depicts the sum secrecy rate versus the number of IRS elements $N$. The performance of both IRS-assisted schemes (IRS-RSMA and IRS-NOMA) improves with larger $N$, as a larger IRS offers greater beamforming resolution in the reflection domain, allowing for more effective signal focusing and nulling. The proposed IRS-RSMA scheme exhibits a steeper growth, indicating a more efficient utilization of the IRS’s capabilities compared to NOMA. The schemes without IRS are, of course, unaffected by $N$.
Optimized UAV Deployment
The PSO algorithm consistently positions the UAV drone closer to the cluster of legitimate users and strategically between the users and the IRS. This positioning minimizes the average path loss to the users while maintaining a strong LoS link to the IRS, enabling the IRS to effectively act as a signal focal point. This intelligent deployment, facilitated by the UAV’s mobility, is a key factor in achieving the high secrecy rates.
Conclusion and Future Work
This paper has investigated the critical problem of physical layer security in IRS-assisted UAV communication systems employing RSMA. We formulated a sum secrecy rate maximization problem by jointly optimizing the UAV drone’s position, transmit precoding, common rate allocation, and IRS phase shifts in the presence of multiple eavesdroppers. To solve this intricate non-convex problem, a novel hierarchical optimization framework was proposed, decomposing it into an inner block for communication resource allocation (solved via alternating optimization with SCA and SDR) and an outer block for UAV positioning (solved via PSO). This approach effectively handles the coupling between the UAV’s mobility and the configurable wireless environment created by the IRS.
Simulation results demonstrate that the proposed algorithm significantly outperforms benchmark schemes, including those using NOMA or operating without IRS assistance. The results validate the synergistic benefits of integrating IRS, RSMA, and UAV mobility: the UAV provides a favorable aerial platform, RSMA manages multi-user interference robustly, and the IRS actively enhances secrecy by reconfiguring the propagation channels. This work provides a valuable algorithmic framework for designing secure, high-throughput aerial networks, with direct relevance to the advancement of China UAV drone applications in future 6G systems.
Future research directions include: 1) Considering discrete phase shifts at the IRS to align with practical hardware constraints; 2) Investigating robust optimization under imperfect channel state information (CSI), especially for the eavesdroppers’ channels; 3) Extending the framework to multi-IRS or multi-UAV drone collaborative scenarios for even greater coverage and security; and 4) Incorporating energy efficiency metrics to design green secure communication protocols for sustainable UAV drone operations.