Unmanned aerial vehicles have become one of the fastest-growing platforms in modern wireless networks, relaying broadband traffic, extending coverage into disaster zones, and serving as flexible aerial base stations for the emerging sixth generation of communications. Yet the very openness that makes drone communications so useful also makes them dangerously exposed. A radio signal broadcast from the sky travels over unobstructed paths, and any adversary with a receiver can attempt to detect that a transmission is taking place at all, or worse, decode its contents. A new study published in Aerospace Systems by Lei Xiong of the Jiangxi Institute of Technology in Nanchang addresses this dual problem head-on, proposing a framework in which intelligent reflective surfaces, deep reinforcement learning, and convex optimization work together to keep drone links both invisible to watchers and unreadable to eavesdroppers.
The two security goals at stake are related but distinct. Covert communication concerns concealment: the transmitting drone wants an eavesdropper’s probability of even detecting that a transmission has occurred to remain close to that of random noise, which imposes strict statistical constraints on how much power can be radiated toward any potential watcher. Secure transmission, by contrast, concerns secrecy once a signal is detected or intercepted, measured through the secrecy capacity, the gap between the data rate achieved by the legitimate user and the rate leaking to the eavesdropper. Most prior work has treated these objectives separately, optimizing a drone’s flight path for covertness in one study and its beamforming for secrecy in another. The new research argues that this separation is a mistake, because the same decisions about trajectory, transmit power, and reflection geometry simultaneously shape both detection risk and data leakage.
The centerpiece of the framework is the intelligent reflective surface, sometimes called a reconfigurable intelligent surface. These are planar arrays of dozens or hundreds of nearly passive elements, each of which can impose a controllable phase shift on an incoming radio wave. By coordinating those phase shifts, the system can steer reflected energy toward a legitimate ground user with constructive alignment, while arranging destructive interference along paths that would otherwise carry the signal toward an eavesdropper. In the drone scenario studied here, the reflective surface acts as a programmable mirror in the environment: the UAV transmits toward it, and the surface reshapes the propagation so that the intended user receives a strong signal while the eavesdropper’s channel is deliberately degraded. Because the elements are almost passive, this signal shaping comes at a fraction of the energy cost of active antennas, which matters enormously for battery-limited aerial platforms.
Formally, the study constructs a joint optimization model that includes the unmanned aerial vehicle, the legitimate ground users, the eavesdropper, and the intelligent reflective surface as coupled decision entities. The variables span the drone’s three-dimensional trajectory, its transmit power over time, and the phase configuration of the reflecting elements. The objective is to maximize the minimum average secrecy capacity experienced by the users, subject to a strict covert constraint that bounds the eavesdropper’s detection performance, together with the physical limits of the drone such as its mobility, energy budget, and communication links. This is a notoriously hard problem: the trajectory variables are continuous and coupled to the discrete-ish phase choices of the surface, the covert constraint is non-convex, and the whole system evolves over time as the drone moves.
To make the problem tractable, the author introduces relaxation variables that represent the minimum average safety capacity of the users, effectively converting a max-min formulation into a form that can be attacked with successive approximation. The solution then combines two complementary techniques. Sequential convex approximation is used to handle the non-convex geometry of the trajectory and beamforming subproblems: at each iteration, the intractable terms are replaced by convex surrogates that are tight around the current operating point, and the resulting problem is solved exactly before the approximation is refined. Deep double Q-networks, a reinforcement learning architecture that reduces the overestimation bias of standard Q-learning by decoupling action selection from action evaluation, learn the end-to-end coordination policy across the joint action space. The hybrid framework thus achieves what the paper describes as end-to-end collaborative optimization, in which the learned policy and the mathematical refinement loop reinforce each other rather than competing.
The reported numerical results are striking. Under the hybrid optimization framework and while satisfying the strict concealment constraints, the proposed method achieves a secrecy capacity of 2.41 megabits per second, meaning the legitimate link enjoys a substantial rate advantage over anything the eavesdropper can extract. Total energy consumption across the mission is only 4.82 kilojoules, which the study reports as significantly lower than all comparison algorithms tested. Perhaps most telling for real deployments, the average energy efficiency of the unmanned aerial vehicle reaches 5.94 megabits per joule, a figure that quantifies how much secure data the platform delivers for every unit of energy it expends. Taken together, these numbers indicate that the framework does not merely trade throughput for stealth; it finds operating points where concealment, secrecy, and endurance are simultaneously respected.
The comparisons against baseline algorithms illuminate why the joint approach wins. Methods that optimize the drone’s path without adapting the reflective surface, or that tune the surface phases without accounting for the covert detection constraint, inevitably leave performance on the table because they ignore the coupling between the decisions. Classical optimization-only baselines struggle with the scale and non-convexity of the joint problem, while pure learning baselines can be sample-inefficient and imprecise near the constraint boundaries. The hybrid design, by embedding the learned policy inside a convex refinement loop, inherits the adaptability of reinforcement learning for handling the dynamic, partially observable aerial environment while retaining the rigor of mathematical optimization where guarantees matter most, namely in keeping the eavesdropper’s detection probability within the covert threshold.
The broader context makes the contribution timely. Recent surveys of sixth-generation architectures describe the integration of autonomous aerial vehicles with intelligent reflecting surfaces as a defining paradigm, and a growing literature applies deep learning to secure UAV-assisted reflective networks, robust designs against cooperative hostile detection, and short-packet covert links with dual drones. What distinguishes the present work is the explicit insistence that concealment and secrecy be optimized as a pair under a single constraint set, with energy efficiency treated as a first-class objective rather than an afterthought. For applications such as battlefield relay, critical infrastructure inspection, and privacy-sensitive urban connectivity, a drone that conserves battery while remaining undetectable and unexploitable is precisely the profile that operators need.
Challenges remain before such systems leave the simulation environment. The covert constraint depends on statistical models of the eavesdropper’s receiver and detector, and real adversaries may behave in ways those models do not capture. The reflective surface must be deployed somewhere in the environment, and its placement, channel estimation accuracy, and hardware imperfections will all influence the achievable gains. Reinforcement learning policies also require careful training and validation before they can be trusted in safety-critical airspace. Still, the study demonstrates that the tools already exist to make aerial wireless links dramatically harder to detect and harder to break, at energy budgets that small drones can actually afford. As intelligent surfaces migrate from laboratory prototypes to building facades and street furniture, the vision of drone networks that whisper instead of shout moves closer to engineering reality, and this research offers a concrete, quantified blueprint for how that whispering can be orchestrated.
Subject of Research: Covert communication and secure transmission of unmanned aerial vehicles assisted by intelligent reflective surfaces
Article Title: Covert communication and secure transmission of unmanned aerial vehicles assisted by intelligent reflective surfaces
Article References: Xiong, L. (2026). Covert communication and secure transmission of unmanned aerial vehicles assisted by intelligent reflective surfaces. Aerospace Systems. https://doi.org/10.1007/s42401-026-00541-7
Image Credits: AI Generated
DOI: 10.1007/s42401-026-00541-7
Keywords: Covert, communication, secure, transmission, unmanned, aerial, vehicles, assisted, intelligent, reflective, surfaces, scientific research
News Source: Denise Maddox. (October 13, 2026). Smart Reflective Surfaces Teach Drones to Hide Their Signals From Eavesdroppers. Scienmag.
