In the era of fifth-generation and beyond wireless networks, unmanned aerial vehicles have become indispensable for a wide range of applications such as emergency rescue, surveillance, and temporary coverage. However, the broadcast nature of wireless channels makes the links involving unmanned aerial vehicles vulnerable to eavesdropping, while the limited on-board energy of unmanned aerial vehicles imposes a critical constraint on their operational endurance. Intelligent reflecting surfaces (IRSs) have emerged as a promising technology to enhance both spectral efficiency and physical layer security by reconfiguring the propagation environment. In this work, I focus on the secure communication energy efficiency of IRS-assisted unmanned aerial vehicle systems. I first consider an ideal hardware scenario and jointly optimize the IRS phase shifts, the trajectory, and the velocity of the unmanned aerial vehicle. To solve the non-convex fractional problem, I apply alternating optimization combined with Dinkelbach’s transform, successive convex approximation (SCA), and manifold optimization. In the second part, I extend the framework to account for transceiver hardware impairments (HIs). The statistical characteristics of HIs are incorporated into the signal model, and a robust design is proposed by jointly optimizing the active beamforming at the unmanned aerial vehicle base station, the passive beamforming at the IRS, and the trajectory and velocity of the unmanned aerial vehicle. Semidefinite programming (SDP) is used for the beamforming subproblems, while SCA is employed for the trajectory subproblem. Simulation results verify that the proposed schemes can significantly improve the secure communication energy efficiency and effectively mitigate the adverse impact of hardware impairments.
1. Introduction
The rapid development of wireless communications has led to an increasing need for flexible and reliable connectivity. Unmanned aerial vehicles offer a unique combination of mobility, altitude, and deployment flexibility, making them attractive for both military and civil applications. Nevertheless, when a unmanned aerial vehicle acts as an aerial base station or relay, the line-of-sight (LoS) dominated channels may expose the transmitted information to malicious eavesdroppers. Physical layer security has therefore become an important paradigm for protecting confidential messages without relying on complex upper-layer encryption. One efficient approach is to exploit the additional degrees of freedom provided by the movement of unmanned aerial vehicles as well as the passive beamforming capability of IRSs.
An IRS is composed of a large number of low-cost passive reflecting elements, each capable of introducing a controllable phase shift to the incident signal. By properly adjusting these phase shifts, the reflected signals can coherently combine with the direct signals at the intended receiver, thereby enhancing the legitimate link. Conversely, the reflected signals can be made to interfere destructively at the eavesdropper, reducing its received signal-to-noise ratio. The integration of IRSs with unmanned aerial vehicles creates a powerful synergy: the IRS can compensate for the lack of direct coverage when the unmanned aerial vehicle cannot be placed close to the ground users, while the unmanned aerial vehicle trajectory can be designed to reduce flight energy consumption and to improve the propagation environment.
Most existing studies on IRS-assisted unmanned aerial vehicle secure communications focus on maximizing the secrecy rate. However, the energy consumption of the unmanned aerial vehicle is a critical issue that affects the practical deployment. A more holistic performance metric is the secure communication energy efficiency, defined as the ratio of the achievable secrecy rate to the total propulsion energy consumption of the unmanned aerial vehicle. This metric balances the conflicting objectives of enhancing security and prolonging the mission duration.
In this paper, I study the secure communication energy efficiency of IRS-assisted unmanned aerial vehicle systems. I first formulate a joint optimization problem involving the IRS phase shifts, the trajectory, and the speed of the unmanned aerial vehicle. Because the problem is highly non-convex and contains a fractional objective, I decompose it into three subproblems and solve them using alternating optimization. For the IRS phase shift design, I adopt a manifold optimization approach that operates directly on the complex circle manifold, thus preserving the constant modulus constraint. For the trajectory and speed design, I employ Dinkelbach’s transform to handle the fractional objective and use SCA to convexify the non-convex constraints. I then extend the framework to a more practical scenario where the transceivers suffer from residual hardware impairments. The presence of HIs introduces additional distortion noise whose power scales with the signal power. This changes the structure of the secrecy rate and requires a reformulation of the optimization problem. I propose a three-stage iterative algorithm that alternatingly updates the UAV beamforming vector, the IRS phase shift matrix, and the trajectory and speed of the unmanned aerial vehicle.
2. System Model and Problem Formulation
2.1 Ideal Hardware Case
I consider a downlink communication system where a single-antenna unmanned aerial vehicle serves a ground legitimate user in the presence of a passive eavesdropper. An IRS with \(M\) reflecting elements is deployed on a building facade to assist the transmission. The unmanned aerial vehicle flies at a fixed altitude \(H_U\) for a total duration \(T\), which is divided into \(N\) time slots of equal length \(\delta_t = T/N\). The horizontal coordinate of the unmanned aerial vehicle at slot \(n\) is denoted by \(\mathbf{q}[n]\), and its velocity is denoted by \(\mathbf{v}[n]\). The trajectory constraints are expressed as:
\[
\begin{aligned}
\mathbf{q}[n] &= \mathbf{q}[n-1] + \delta_t \mathbf{v}[n], \quad n=2,\dots,N, \\
\mathbf{q}[1] &= \mathbf{q}_I + \delta_t \mathbf{v}[1], \\
\mathbf{q}[N] &= \mathbf{q}_F, \\
\|\mathbf{v}[n]\| &\le V_{\max},
\end{aligned}
\]
where \(\mathbf{q}_I\) and \(\mathbf{q}_F\) are the initial and final horizontal positions, respectively. The legitimate user and the eavesdropper are located at \(\mathbf{w}_L = [x_L, y_L]^T\) and \(\mathbf{w}_E = [x_E, y_E]^T\), respectively. The IRS position is denoted by \(\mathbf{w}_R = [x_R, y_R]^T\). The channel between the unmanned aerial vehicle and the IRS is modeled as a LoS link, while the channels involving ground users are modeled as Rician fading. The received signal at user \(i \in \{L,E\}\) at slot \(n\) is
\[
y_i[n] = \left( h_{Ui}[n] + \mathbf{h}_{Ri}^H \mathbf{\Theta}[n] \mathbf{h}_{UR}[n] \right) s[n] + n_i[n],
\]
where \(h_{Ui}[n]\) is the direct channel between the unmanned aerial vehicle and user \(i\), \(\mathbf{h}_{Ri}\) is the channel between the IRS and user \(i\), \(\mathbf{h}_{UR}[n]\) is the channel between the unmanned aerial vehicle and the IRS, and \(\mathbf{\Theta}[n] = \mathrm{diag}(e^{j\theta_1[n]}, \dots, e^{j\theta_M[n]})\) is the phase shift matrix of the IRS. The noise \(n_i[n]\) is zero-mean complex Gaussian with variance \(\sigma_i^2\).
The achievable rate of user \(i\) is given by
\[
R_i[n] = B \log_2 \left( 1 + \frac{P | h_{Ui}[n] + \mathbf{h}_{Ri}^H \mathbf{\Theta}[n] \mathbf{h}_{UR}[n]|^2}{\sigma_i^2} \right),
\]
where \(B\) is the bandwidth and \(P\) is the transmit power of the unmanned aerial vehicle. The secrecy rate at slot \(n\) is
\[
R_s[n] = \left[ R_L[n] – R_E[n] \right]^+,
\]
with \([x]^+ = \max(x,0)\). The total propulsion energy consumption of the unmanned aerial vehicle is modeled as
\[
E_{\mathrm{prop}} = \sum_{n=1}^{N} \delta_t \left( c_1 \|\mathbf{v}[n]\|^3 + \frac{c_2}{\|\mathbf{v}[n]\|} \right),
\]
where \(c_1\) and \(c_2\) are parameters depending on the unmanned aerial vehicle weight, wing area, and air density. The secure communication energy efficiency is defined as
\[
\zeta = \frac{\sum_{n=1}^{N} R_s[n]}{E_{\mathrm{prop}}}.
\]
The joint optimization problem in the ideal hardware case is formulated as
\[
\begin{aligned}
\max_{\mathbf{\Theta}, \mathbf{Q}, \mathbf{V}} \quad & \zeta \\
\text{s.t.} \quad & |e^{j\theta_m[n]}| = 1, \quad \forall m,n, \\
& (1), (2), (3), (4).
\end{aligned}
\]
2.2 Hardware Impairments Case
In practice, the transceivers of the unmanned aerial vehicle and the ground users suffer from residual hardware impairments such as phase noise, power amplifier nonlinearities, and I/Q imbalance. These impairments are typically modeled as additive distortion noises. I assume a multiple-antenna unmanned aerial vehicle base station with \(N_t\) antennas. At slot \(n\), the transmitted signal is \(\mathbf{w}[n] s[n]\), where \(\mathbf{w}[n] \in \mathbb{C}^{N_t \times 1}\) is the beamforming vector. The received signal at user \(i\) is
\[
y_i[n] = \left( \mathbf{h}_{Ui}^H[n] + \mathbf{h}_{Ri}^H \mathbf{\Theta}[n] \mathbf{H}_{UR}[n] \right) \mathbf{w}[n] s[n] + \eta_{T}[n] + \eta_{R,i}[n] + n_i[n],
\]
where \(\mathbf{h}_{Ui}[n] \in \mathbb{C}^{N_t \times 1}\) is the channel between the unmanned aerial vehicle and user \(i\), \(\mathbf{H}_{UR}[n] \in \mathbb{C}^{M \times N_t}\) is the channel between the unmanned aerial vehicle antenna array and the IRS, and \(\mathbf{h}_{Ri} \in \mathbb{C}^{M \times 1}\) is the channel between the IRS and user \(i\). The distortion noises are distributed as
\[
\eta_{T}[n] \sim \mathcal{CN}\left(0, \mu_t \mathrm{diag}(\mathbf{w}[n] \mathbf{w}^H[n]) \right),
\]
\[
\eta_{R,i}[n] \sim \mathcal{CN}\left(0, \mu_{r,i} \mathbb{E}\left\{ |\tilde{y}_i[n]|^2 \right\} \right),
\]
where \(\mu_t\) and \(\mu_{r,i}\) are non-negative constants characterizing the severity of the hardware impairments, and \(\tilde{y}_i[n]\) is the undistorted received signal. Based on the statistical properties of these distortion noises, the effective signal-to-noise-plus-distortion ratio at user \(i\) is derived as
\[
\gamma_i[n] = \frac{ \mathbf{w}^H[n] \mathbf{G}_i[n] \mathbf{w}[n] }{ \mu_t \mathbf{w}^H[n] \mathrm{diag}(\mathbf{G}_i[n]) \mathbf{w}[n] + \mu_{r,i} \mathbf{w}^H[n] \mathbf{G}_i[n] \mathbf{w}[n] + \sigma_i^2 },
\]
where \(\mathbf{G}_i[n] = (\mathbf{h}_{Ui}[n] + \mathbf{H}_{UR}^H[n] \mathbf{\Theta}^H[n] \mathbf{h}_{Ri})(\mathbf{h}_{Ui}[n] + \mathbf{H}_{UR}^H[n] \mathbf{\Theta}^H[n] \mathbf{h}_{Ri})^H\). The rate of user \(i\) is \(R_i[n] = B \log_2(1 + \gamma_i[n])\). The optimization problem with hardware impairments is formulated as
\[
\begin{aligned}
\max_{\mathbf{W}, \mathbf{\Theta}, \mathbf{Q}, \mathbf{V}} \quad & \frac{\sum_n [R_L[n] – R_E[n]]^+}{E_{\mathrm{prop}}} \\
\text{s.t.} \quad & \|\mathbf{w}[n]\|^2 \le P_{\max}, \quad \forall n, \\
& |e^{j\theta_m[n]}| = 1, \quad \forall m,n, \\
& (1)-(4).
\end{aligned}
\]
3. Proposed Optimization Framework
3.1 Alternating Optimization Structure
The formulated problems are non-convex and involve coupled variables. To make them tractable, I adopt an alternating optimization procedure. At each iteration, I fix the phase shifts and optimize the trajectory and velocity, then fix the trajectory and velocity to optimize the phase shifts. For the hardware impairment case, I also insert an additional step to optimize the beamforming vector. The overall algorithm is guaranteed to converge to a stationary point because each subproblem is solved to optimality and the objective value is non-decreasing after each update.
A summary of the alternating optimization framework is given in Table 1.
| Stage | Fixed variables | Optimized variables | Method |
|---|---|---|---|
| 1 | \(\mathbf{\Theta}\) | \(\mathbf{Q}, \mathbf{V}\) | Dinkelbach + SCA |
| 2 | \(\mathbf{Q}, \mathbf{V}\) | \(\mathbf{\Theta}\) | Manifold optimization (ideal) / SDP (with HIs) |
| 3 (HIs only) | \(\mathbf{Q}, \mathbf{V}, \mathbf{\Theta}\) | \(\mathbf{W}\) | SDP |
3.2 Trajectory and Velocity Optimization
For fixed \(\mathbf{\Theta}\), the trajectory and velocity subproblem is a fractional program. I introduce slack variables to handle the non-convex terms. The energy consumption is transformed using a slack variable \(\tau[n]\) with the constraint \(\tau[n] \le \|\mathbf{v}[n]\|\). For the secrecy rate, I introduce variables to separate the distance terms. After applying first-order Taylor expansions, the subproblem becomes convex. Dinkelbach’s method is then used to handle the fraction by iteratively solving
\[
\max_{\mathbf{Q}, \mathbf{V}, \mathbf{t}, \mathbf{s}, \mathbf{\eta}} \quad \sum_{n=1}^N \left( \hat{R}_s[n] – \lambda^{(r)} \hat{E}_{\mathrm{prop}}[n] \right),
\]
where \(\lambda^{(r)}\) is updated as
\[
\lambda^{(r)} = \frac{\sum_n \hat{R}_s^{(r)}[n]}{\hat{E}_{\mathrm{prop}}^{(r)}}.
\]
The convexified constraints include
\[
\|\mathbf{v}[n]\|^2 \ge \|\mathbf{v}_0[n]\|^2 + 2 \mathbf{v}_0^T[n] (\mathbf{v}[n] – \mathbf{v}_0[n]),
\]
and similar linearization for the distance-related terms. The final subproblem is solved by CVX.
3.3 IRS Phase Shift Optimization: Ideal Hardware Case
When the trajectory and velocity are fixed, the numerator of the objective is the only part that depends on \(\mathbf{\Theta}\). I define the vector \(\mathbf{z}[n] = [e^{j\theta_1[n]}, \dots, e^{j\theta_M[n]}, 1]^T\). Then the channel gain can be written as \(\mathbf{z}^H[n] \mathbf{H}_i[n] \mathbf{z}[n]\). The subproblem becomes
\[
\min_{\mathbf{z}[n] \in \mathcal{M}} \quad \sum_{n=1}^N f_n(\mathbf{z}[n]),
\]
where \(\mathcal{M} = \{ \mathbf{z} : |z_1| = \dots = |z_{M+1}| = 1 \}\) is a complex circle manifold, and
\[
f_n(\mathbf{z}[n]) = – \sum_{i \in \{L,E\}} \alpha_i \log_2\left(1 + \frac{P}{c_i} \mathbf{z}^H[n] \mathbf{H}_i[n] \mathbf{z}[n] \right).
\]
I solve this problem using manifold optimization. The Riemannian gradient is obtained by projecting the Euclidean gradient onto the tangent space. A conjugate gradient direction is computed using the Polak-Ribiere parameter, and a retraction step maps the updated point back to the manifold. The algorithm is summarized in Table 2.
| Step | Operation |
|---|---|
| 1 | Initialize \(\mathbf{z}_0[n]\), set \(k=0\) |
| 2 | Compute Euclidean gradient \(\nabla f(\mathbf{z}_k[n])\) |
| 3 | Compute Riemannian gradient \(\mathrm{grad}f = \nabla f – \mathrm{Re}\{\nabla f \odot \mathbf{z}_k^*\} \odot \mathbf{z}_k\) |
| 4 | Update search direction \(\mathbf{s}_k = -\mathrm{grad}f + \rho_k \mathbf{s}_{k-1}\) |
| 5 | Armijo line search for step size \(\alpha_k\) |
| 6 | \(\tilde{\mathbf{z}}_k = \mathbf{z}_k + \alpha_k \mathbf{s}_k\) |
| 7 | Retraction: \(\mathbf{z}_{k+1} = \tilde{\mathbf{z}}_k / |\tilde{\mathbf{z}}_k|\) |
| 8 | Convergence check; otherwise set \(k = k+1\) |
3.4 IRS Phase Shift and Beamforming Optimization with Hardware Impairments
In the presence of HIs, the IRS phase shift subproblem and the UAV beamforming subproblem are optimized using semidefinite relaxation. For the beamforming design, I define \(\mathbf{W}[n] = \mathbf{w}[n] \mathbf{w}^H[n]\). The subproblem becomes an SDP after applying SCA to handle the log functions. Similarly, for the IRS phase shifts, I define \(\mathbf{P}[n] = \mathbf{p}[n] \mathbf{p}^H[n]\) with \(\mathbf{p}[n] = [e^{j\theta_1[n]}, \dots, e^{j\theta_M[n]}, 1]^T\). The rank-one constraint is dropped using the SDR technique, and the resulting SDP is solved efficiently by CVX. The overall algorithm for the hardware impairment case is summarized in Table 3.
| Step | Operation |
|---|---|
| 1 | Initialize \(\mathbf{Q}^{(0)}, \mathbf{V}^{(0)}, \mathbf{\Theta}^{(0)}, \mathbf{W}^{(0)}\), set \(j=0\) |
| 2 | Repeat |
| 3 | Given \(\mathbf{\Theta}^{(j)}, \mathbf{W}^{(j)}\), solve trajectory subproblem to obtain \(\mathbf{Q}^{(j+1)}, \mathbf{V}^{(j+1)}\) |
| 4 | Given \(\mathbf{Q}^{(j+1)}, \mathbf{V}^{(j+1)}, \mathbf{\Theta}^{(j)}\), solve SDP for \(\mathbf{W}^{(j+1)}\) |
| 5 | Given \(\mathbf{Q}^{(j+1)}, \mathbf{V}^{(j+1)}, \mathbf{W}^{(j+1)}\), solve SDP for \(\mathbf{\Theta}^{(j+1)}\) |
| 6 | Update the objective value |
| 7 | Set \(j = j+1\) until convergence |
4. Simulation Results and Discussion
4.1 Parameter Settings
The simulation parameters used for both scenarios are listed in Table 4. The unmanned aerial vehicle flies at a constant altitude of \(100\) m in the ideal case and \(80\) m in the hardware impairment case. The IRS is placed at a height of \(40\) m with \(M = 64\) elements in the ideal case and \(M = 32\) elements in the hardware impairment case. The maximum speed of the unmanned aerial vehicle is \(V_{\max} = 30\) m/s. The bandwidth is \(B = 1\) MHz and the noise power is \(\sigma_L^2 = \sigma_E^2 = -80\) dBm. For the hardware impairment case, the distortion coefficients are \(\mu_t = \mu_{r,l} = \mu_{r,e} = 0.1\).
| Parameter | Ideal Case | Hardware Impairments Case |
|---|---|---|
| Path loss exponents | \(\alpha=2, K=2.5\) | \(\alpha_{ur}=2, \alpha_u=3.5, \alpha_r=3.5\) |
| UAV altitude | \(100\) m | \(80\) m |
| IRS elements | \(M=64\) | \(M=32\) |
| UAV max speed | \(V_{\max}=30\) m/s | \(V_{\max}=30\) m/s |
| Transmit power | \(P=20\) dBm | \(P_{\max}=20\) dBm |
| Channel power gain | \(\rho=-30\) dB | \(\rho=-30\) dB |
| UAV antennas | \(1\) | \(N_t=16\) |
| IRS position | \((0,40)\) m | \((0,40)\) m |
| Legitimate user | \((0,120)\) m | \((0,120)\) m |
| Eavesdropper | \((150,150)\) m | \((150,150)\) m |
| UAV start | \((-300,30)\) m | \((-300,30)\) m |
| UAV end | \((300,30)\) m | \((300,30)\) m |
4.2 UAV Trajectory and Speed
Figure 1 demonstrates the optimized trajectory of the unmanned aerial vehicle in the ideal hardware case. The unmanned aerial vehicle initially moves toward the legitimate user in order to improve the direct link quality. However, instead of flying directly above the legitimate user, the trajectory is biased toward the IRS location. This is because the IRS provides an additional virtual link that can significantly boost the legitimate signal strength without requiring the unmanned aerial vehicle to consume extra energy to hover near the user. Once the unmanned aerial vehicle reaches a point where the legitimate channel quality is sufficiently high, it accelerates toward the final destination. The corresponding speed profile shows that the unmanned aerial vehicle reduces its speed in the middle of the mission to maintain a high signal quality, and then increases the speed near the end to reduce the total flight time.
In the hardware impairment case, the trajectory follows a similar trend. The presence of HIs does not change the qualitative shape of the trajectory, but the distances become slightly larger because the effective signal power is reduced by the distortion noise. The optimized trajectory in all cases respects the maximum speed constraint and the boundary constraints.
4.3 Energy Efficiency vs. Flight Time
Figure 2 shows the secure communication energy efficiency as a function of the total flight time \(T\) in the ideal hardware case. For all schemes, the energy efficiency decreases as \(T\) increases. This is due to the fact that the total energy consumption grows linearly with the flight time, while the secrecy rate saturates after the unmanned aerial vehicle has traveled far enough from the eavesdropper. The proposed manifold-based alternating algorithm achieves a significantly higher energy efficiency than the baseline that only optimizes the secrecy rate (without considering the energy). When the IRS is absent, the energy efficiency is much lower because the unmanned aerial vehicle must fly closer to the legitimate user to achieve a comparable secrecy rate, thus consuming more energy.
Compared with the SDR-based approach for the IRS phase shifts, the manifold optimization method yields almost identical energy efficiency but with lower computational complexity. This makes the proposed algorithm suitable for real-time implementation on energy-constrained unmanned aerial vehicles.
4.4 Effect of Maximum Speed
Figure 3 plots the energy efficiency versus the maximum speed \(V_{\max}\). As the maximum speed increases, the feasible region of the trajectory and speed optimization expands, allowing the unmanned aerial vehicle to fly faster when beneficial. For the proposed energy-efficiency-oriented scheme, the objective value first increases and then saturates. In contrast, the secrecy-rate-oriented baseline shows a decreasing energy efficiency as \(V_{\max}\) grows, because the baseline does not account for the propulsion energy and simply increases the speed, leading to a cubic growth in energy consumption.
4.5 Convergence Behavior
Table 5 presents the convergence behavior of the proposed alternating algorithm for different flight times. The algorithm converges within fewer than 15 iterations for all tested scenarios. The convergence is monotonic because each subproblem is solved to optimality and the objective value is non-decreasing. The computational time per iteration is moderate, which is acceptable for offline trajectory planning.
| Flight time \(T\) | Iterations to converge | Objective value at convergence (bit/s/Hz/J) |
|---|---|---|
| 40 s | 12 | 2.31 |
| 50 s | 10 | 1.98 |
| 60 s | 9 | 1.72 |
4.6 Energy Efficiency vs. Maximum Transmit Power
In the hardware impairment case, the secure communication energy efficiency is plotted versus the maximum transmit power \(P_{\max}\) in Figure 4. The energy efficiency increases with \(P_{\max}\), but the growth rate diminishes as the distortion noise power also increases with the transmit power. This indicates that the hardware impairments set a fundamental limit on the achievable energy efficiency. The proposed scheme, which explicitly takes the HIs into account during the beamforming design, outperforms the scheme that ignores the HIs. When the IRS is not deployed, the energy efficiency is the lowest, demonstrating the benefit of the IRS in combating both the path loss and the distortion noise.
4.7 Impact of Hardware Impairment Severity
Figure 5 shows the energy efficiency versus the HIs coefficient \(\mu\). As \(\mu\) increases, the distortion noise power grows, leading to a lower secrecy rate and thus a lower energy efficiency. The proposed HIs-aware design achieves a higher energy efficiency than the design that assumes ideal hardware, because the HIs-aware design optimizes the beamformer to balance the signal power and the distortion. The performance gap becomes larger as \(\mu\) increases, emphasizing the importance of incorporating HIs into the design.
4.8 Effect of the Number of UAV Antennas
Figure 6 depicts the energy efficiency versus the number of antennas \(N_t\) on the unmanned aerial vehicle. A larger \(N_t\) provides more spatial degrees of freedom, enabling the unmanned aerial vehicle to direct the signal more precisely toward the legitimate user and away from the eavesdropper. Thus, the energy efficiency improves monotonically with \(N_t\). The proposed scheme again outperforms all baseline schemes, especially when \(N_t\) is large.
5. Conclusions
In this work, I have investigated the secure communication energy efficiency of IRS-assisted unmanned aerial vehicle systems under both ideal transceiver hardware and transceiver hardware impairments. I formulated joint optimization problems involving the IRS phase shifts, the trajectory and velocity of the unmanned aerial vehicle, and, for the hardware impairment case, the active beamforming vector. The problems are highly non-convex and were solved by alternating optimization with carefully tailored techniques. The trajectory subproblem was converted into a convex problem using SCA and Dinkelbach’s transform. The IRS phase shift subproblem was solved by manifold optimization in the ideal case and by semidefinite programming in the hardware impairment case. The beamforming subproblem was also solved by SDP. Simulation results demonstrated that the proposed schemes can significantly improve the secure communication energy efficiency compared with existing baselines. The manifold optimization approach achieves nearly the same performance as SDR but with lower complexity. The HIs-aware design effectively suppresses the adverse impact of hardware impairments, and the energy efficiency improves as the number of antennas increases. Future work can extend the framework to multi-unmanned aerial vehicle networks, imperfect channel state information, and discrete IRS phase shifts.