Joint UAV trajectory and offloading optimization with robust secrecy for intelligent mining Authors: Abdulbasit A Darem, Khalid Almalki, Abdulrahman Alzahrani, Khalid N R Alharbi, Asma A Alhashmi, Laith A Darem Published: 2026 ABSTRACT AbstractIntelligent mining demands real-time processing of UAV sensor streams under latency, safety, and energy constraints. We study a dual-edge architecture in which a ground base station (BS) and an aerial edge server (AES) collaboratively serve aerial users (FDMA), while a protective jammer UAV adapts its trajectory subject to speed/acceleration limits and rotary-wing propulsion. The system models 3GPP A2G channels (probabilistic LoS/NLoS and state-conditioned fading) and passive ground-eavesdroppers with bounded location uncertainty. We formulate a robust energy-efficiency maximization that epigraphs worst-case eavesdropper rates, introduces secrecy-QoS slack variables for feasibility, and enforces slot-level task causality. The fractional objective is handled via Dinkelbach, and three-block BCD solves the problem with conservative SCA surrogates; a micro-AO resolves the bi-convex throughput epigraph in the radio block, and the trajectory block uses affine secrecy bounds plus an SOC treatment of induced power in the rotary-wing model. Under 3GPP Urban Macro calibration, the proposed scheme attains 60.5 kbits/J at U=10, exceeding STRO by 22.2% and SHJ by 116.1%. With eavesdropper uncertainty m, energy efficiency degrades only 13.1%, whereas non-robust CENR collapses to 12.0 kbits/J (76.9% drop). The optimized jammer path is 1250 m (vs. 1131 m straight line; +10.5%) to secure stronger jamming geometry; propulsion dominates the energy budget at 4800 J (86.8% of total), while SLT despite saving 6.3% propulsion energy, incurs a 730.9% increase in RF jamming to 2742 J. The algorithm runs in 1.95 s/iteration on average (trajectory-frozen: 0.98 s/iteration), converges within 5–7 iterations, and captures 85% of its total gain in the first 3 iterations–validating near-real-time feasibility for secure, energy-efficient offloading in mining operations.Keywords: Aerial-edge computing, UAV trajectory optimization, Robust physical-layer secrecy, Cooperative jamming, Dinkelbach transform, Block coordinate descent, Successive convex approximation, Rotary-wing propulsion, FDMA, Intelligent miningSubject terms: Engineering, Mathematics and computing Introduction The convergence of artificial intelligence, Mobile Edge Computing (MEC), and Unmanned Aerial Vehicle (UAV) is rapidly transforming industrial operations across multiple sectors1,2. Intelligent mining stands out as a particularly demanding application due to its reliance on continuous, high rate sensing and stringent latency requirements for safety critical decision making3. Aerial users in these environments generate computationally intensive tasks4, such as real-time geological analysis, hazard detection, and equipment monitoring, that must be executed under realistic Air-to-Ground (A2G) propagation conditions and limited onboard processing capabilities5,6. Specifically, these tasks involve heavy payloads such as high-resolution video streams, 3D LiDAR point clouds, and multispectral telemetry that require massive processing power (e.g., over 1000 Central Processing Unit (CPU) cycles/bit). Because mining UAVs have severely constrained batteries, and distant cloud servers violate the strict sub-second latency (Δ = 1 s) required for emergency hazard alerts, real-time edge offloading is functionally mandatory. These operational demands introduce interconnected challenges that existing frameworks do not fully address. The radio environment follows complex 3rd Generation Partnership Project (3GPP)- specified propagation characteristics, with probabilistic Line-of-Sight (LoS) and state-conditioned fading7,8. Furthermore, the wireless medium is vulnerable to passive ground eavesdroppers whose locations are only imperfectly known, directly threatening the confidentiality of offloaded operational data9,10. Simultaneously, UAV platforms must meticulously balance communication, computation, and mobility under strict aerodynamic propulsion limits and kinematic constraints11,12. While aerial edge computing offers a promising paradigm by offloading tasks to nearby ground-based and Aerial Edge Server (AES)13,14, no unified framework currently exists that jointly optimizes radio resource allocation, computation scheduling, UAV trajectory design15, and security against location uncertain eavesdroppers in a manner explicitly tailored to intelligent mining deployments. Research questions RQ1: Joint optimization. How can communication, computation, and UAV mobility be jointly optimized to maximize system Energy Efficiency (EE) under latency and processing limits? RQ2: Robust secrecy. How can secure offloading be guaranteed against worst-case eavesdroppers with location uncertainty while maintaining optimization feasibility? RQ3: Jammer trajectory design. How should a protective jammer UAV shape its trajectory to trade propulsion energy for improved jamming geometry and secrecy? RQ4: Efficient solution method. How can the resulting highly non-convex fractional program coupling secrecy, power control, Frequency Division Multiple Access (FDMA) bandwidth, CPU scheduling, and UAV kinematics—be solved efficiently enough for near-real-time deployment? RQ5: Architectural advantage. Under what conditions and workload regimes does a dual-edge architecture with both a ground Base Station (BS) and an AES outperform single-edge baselines in intelligent mining? To answer these research questions, we develop a comprehensive optimization framework that captures the nonlinear interdependencies between secrecy, propulsion, radio resource allocation, and computation scheduling5,9. The proposed approach integrates a worst-case robust secrecy model with bounded eavesdropper uncertainty, enforces strict task causality constraints, and incorporates realistic UAV kinematics alongside rotary wing propulsion dynamics10,11. Through a novel combination of the Dinkelbach transformation, Block Coordinate Descent (BCD), and Successive Convex Approximation (SCA) with embedded micro Alternating Optimization (AO)14,16, we transform the intractable original problem into a sequence of efficiently solvable convex subproblems. Extensive simulations calibrated with 3GPP Urban Macro (UMa) parameters demonstrate that our framework achieves significant EE gains over multiple baselines, maintains robust performance under severe location uncertainty, and converges rapidly enough for near real-time implementation in dynamic mining environments17,18. The detailed literature review and technical contributions of this work are presented in the following sections. Related work The deployment of UAV in MEC networks has attracted significant research attention. Due to the inherent complexities of these networks, current literature primarily focuses on three intersecting domains: trajectory optimization and resource allocation, learning-based offloading frameworks, and Physical Layer Security (PLS) under eavesdropper uncertainty. Trajectory optimization and task offloading in UAV-MEC The joint optimization of UAV trajectories and computational offloading is fundamental to minimizing system latency and energy consumption. Early works formulated alternative optimization techniques for bit allocation and basic trajectory planning5. To handle broader spatial distributions, optimal transport theory19 and BCD paired with SCA18,20 have been widely applied to large-scale networks. Beyond traditional latency metrics, ensuring information freshness has driven the adoption of Age of Information (AoI) scheduling, which jointly optimizes UAV trajectories and localization accuracy using SCA-based methods21. Furthermore, for remote data collection lacking terrestrial infrastructure, satellite-assisted UAV frameworks have been formulated to maximize system energy efficiency by carefully balancing communication power allocation and trajectory tracking control22. Recent advances expand this paradigm by incorporating task caching and migration via Lyapunov optimization frameworks6, and adopting novel “computing while navigating” schemes leveraging particle swarm optimization to parallelize task execution12. Furthermore, in emergency scenarios, the introduction of Location Priority Levels (LPL) has enabled highly responsive bi-objective optimization for minimization of makespan and energy3. To adapt to highly dynamic and non-convex MEC environments, researchers have increasingly adopted Deep Reinforcement Learning (DRL). Standard DRL algorithms and Double Deep Q-learning (D-DQN) have been effectively utilized to schedule offloading ratios and design Three-Dimensional (3D) trajectories11,23, while Deep Deterministic Policy Gradient (TD3) methods are applied to manage complex Directed Acyclic Graph (DAG) task dependencies1. In hierarchical and multi-UAV edge systems, Optimistic Actor-Critic13 and multi-agent DRL17 networks address computation and communication uncertainties. To manage conflicting system objectives, evolutionary multi-objective reinforcement learning24–26 and game-theoretic models, such as Stackelberg games, have been proposed to optimize UAV utilities and dynamic pricing strategies2,27. As these highly dynamic UAV deployments scale to support the emerging low-altitude economy, researchers are increasingly proposing communication and control co-design strategies within three-layered resilient architectures to actively combat pervasive coverage holes, spectrum scarcity, and volatile wireless links28. Physical layer security in the UAV communications Due to the dominant LoS nature of the air-to-ground channels; the PLS has emerged as a crucial defense mechanism for the UAV networks7. The PLS has been analyzed under the various UAV mobility models; including the random waypoint model29, and extended to the integrated satellite-UAV networks30 as well as the authentication systems managing the jamming uncertainties31. To maximize the secrecy capacity; the modern PLS frameworks integrate the emerging physical layer technologies. These include the Intelligent Reflective Surfaces (IRS) to reconfigure the propagation environments32; the Non-Orthogonal Multiple Access (NOMA) for the Ultra-Reliable Low-Latency Communication (uRLLC)8,33,34, and the Multiple-Input Multiple-Output (MIMO) structures utilizing the transmit antenna selection35. Additionally; the movable antenna technologies36 and the Integrated Sensing and Communication (ISAC) architectures37 have been recently exploited to counter the eavesdroppers lacking the exact location information. Secure UAV-MEC and the robust anti-eavesdropping The intersection of the UAV-assisted MEC and the PLS introduces the severe optimization challenges; particularly when securing the computational offloading links against the malicious eavesdroppers. The cooperative jamming is a prevalent strategy wherein the base stations9,16 or the dedicated multi-role UAVs acting as the jammers14,38 emit the artificial noise (AN) to degrade the eavesdropper’s channel. Crucially; in the many practical scenarios, the exact locations of these eavesdroppers are unknown. Addressing this spatial uncertainty; the robust optimization formulations leveraging the worst-case bounds have been developed for the covert communications in the vehicular networks39,40; the multi-layer aerial-offshore networks utilizing the mobile jammer UAVs10, and the LEO satellite beamforming assisted by the UAV AN41. Despite these significant advancements; the existing literature lacks a unified analytical framework that jointly tackles the dynamic dual-edge MEC resource allocation; the robust worst-case secrecy constraints against the location-uncertain eavesdroppers, and the realistic aerodynamic UAV propulsion models. This paper bridges this gap by proposing a comprehensively coupled Dinkelbach-BCD optimization framework tailored for the highly secure; computation-intensive networks. Motivation As comprehensively detailed in Table 1, the existing literature exhibits several critical limitations that preclude the direct deployment in the highly secure and energy-constrained intelligent mining environments. First, the vast majority of the current MEC studies5,13,19 rely on the simplified kinetic models that fail to capture the highly nonlinear aerodynamic power consumption of the rotary wing UAVs. This fundamental oversight heavily skews the EE optimizations; because hovering and the forward flight require vastly different mechanical power expenditures. Second, while the PLS has gained significant traction; the prior frameworks9,14,16 overwhelmingly assume the perfect knowledge of the passive eavesdropper locations. In the practical mining sites, adversaries deliberately conceal their presence; making the certainty-equivalent network designs highly susceptible to the catastrophic information leaks. Table 1Comprehensive comparison of the proposed framework with existing literature.ReferenceNetwork architecture & task offloadingCoupled resources optimizedUAV propulsion physics modelSecurity mechanismEavesdropper knowledgeSolution approach5,19Single Edge BinaryTrajectory, Offload DecisionSimplified KineticNoneN/AAlternating Optimization6,13Hierarchical Aerial MECTrajectory, Caching, BandwidthSimplified KineticNoneN/ADRL, Lyapunov17Multi UAV MECTrajectory, CPU PartitionSimplified KineticNoneChannel UncertaintyMulti Agent DRL9,14Single Edge PartialTrajectory, Power, CPUSimplified KineticUAV JammingPerfect Location InfoBCD, SCA, ADMM16MEC Aided UAVTrajectory, Transmit PowerSimplified KineticGround Base NoisePerfect Location InfoBCD, SCA39,40Vehicular Aerial DownlinkTrajectory, PowerSimplified KineticCovert RateBounded UncertaintySCA, Deep Learning10Aerial Offshore MECTrajectory, Power, OffloadingSimplified KineticMobile JammerBounded UncertaintyCross Entropy, SCAProposed WorkDual Edge (BS + AES)3D Trajectory, CPU, Power,Nonlinear Rotary WingCooperativeWorst Case BoundedDinkelbach, BCD,Dynamic SplittingBandwidth, Splitting RatiosAerodynamicsUAV JammingSpatial UncertaintyEmbedded Micro SCANote: Simplified kinetic models evaluate flying energy merely proportional to velocity squared, completely ignoring the massive aerodynamic power required for hovering and blade profile drag captured in realistic rotary wing models. Comprehensive comparison of the proposed framework with existing literature. Note: Simplified kinetic models evaluate flying energy merely proportional to velocity squared, completely ignoring the massive aerodynamic power required for hovering and blade profile drag captured in realistic rotary wing models. Furthermore, the limited studies that address the spatial uncertainty10,39,40 typically isolate the communication from the computation; thereby failing to optimize the dual-edge resource coupling required for the intensive task offloading. The deep learning approaches11,23,25, while popular, often lack the strict mathematical convergence guarantees required for the safety critical industrial networks. The distinct absence of a unified mathematical framework that rigorously bounds the worst case secrecy while simultaneously optimizing the fractional EE under the strict rotary wing physics; serves as the primary motivation for this work. To address these critical technical gaps, we propose a comprehensively coupled optimization framework with the following primary contributions; First Integrated Framework for Secure Aerial-Edge Computing in Mining: We propose a novel dual-edge architecture that jointly optimizes the task offloading, the resource allocation, and the trajectory of a protective UAV equipped with a jammer in the intelligent mining environment. Unlike the prior works that treat these elements in isolation; our framework uniquely integrates a 3GPP-compliant A2G channel model (with the probabilistic LoS and the state-conditioned fading) with the physical kinematics of a jammer UAV and a worst-case secrecy model; capturing the complex interdependencies of the real-world mining operations. First Integrated Framework for Secure Aerial-Edge Computing in Mining: We propose a novel dual-edge architecture that jointly optimizes the task offloading, the resource allocation, and the trajectory of a protective UAV equipped with a jammer in the intelligent mining environment. Unlike the prior works that treat these elements in isolation; our framework uniquely integrates a 3GPP-compliant A2G channel model (with the probabilistic LoS and the state-conditioned fading) with the physical kinematics of a jammer UAV and a worst-case secrecy model; capturing the complex interdependencies of the real-world mining operations. A Robust and Tractable Secrecy Formulation with Feasibility Guarantees: We introduce a new robust secrecy model that accounts for the bounded uncertainty in the eavesdropper locations. By epigraphing the worst-case eavesdropper rate and linearizing it via the SCA; we create a convex and safe approximation of the secrecy constraint. Crucially, we incorporate the non-negative slack variables to guarantee the optimization feasibility under the extreme uncertainty; a practical safeguard absent in the most secure offloading studies. This provides a reliable foundation for the secrecy-aware decision-making despite the imperfect location intelligence. A Robust and Tractable Secrecy Formulation with Feasibility Guarantees: We introduce a new robust secrecy model that accounts for the bounded uncertainty in the eavesdropper locations. By epigraphing the worst-case eavesdropper rate and linearizing it via the SCA; we create a convex and safe approximation of the secrecy constraint. Crucially, we incorporate the non-negative slack variables to guarantee the optimization feasibility under the extreme uncertainty; a practical safeguard absent in the most secure offloading studies. This provides a reliable foundation for the secrecy-aware decision-making despite the imperfect location intelligence. A Novel Dinkelbach-BCD-SCA Master Algorithm with Micro-AO: To solve the complex Mixed-Integer Nonlinear Programming (MINLP) problem; we develop a custom solution. We employ a Dinkelbach transform to handle the fractional EE objective and a three-block BCD to decompose the problem; A key innovation is the two-sweep micro-AO embedded within the radio resource block; which efficiently resolves the bi-convexity between the bandwidth allocation and the transmit power/offloading decisions. The complete algorithm is proven to converge monotonically to a stationary point; offering a mathematically sound and efficient path to a high-quality solution. A Novel Dinkelbach-BCD-SCA Master Algorithm with Micro-AO: To solve the complex Mixed-Integer Nonlinear Programming (MINLP) problem; we develop a custom solution. We employ a Dinkelbach transform to handle the fractional EE objective and a three-block BCD to decompose the problem; A key innovation is the two-sweep micro-AO embedded within the radio resource block; which efficiently resolves the bi-convexity between the bandwidth allocation and the transmit power/offloading decisions. The complete algorithm is proven to converge monotonically to a stationary point; offering a mathematically sound and efficient path to a high-quality solution. Aerodynamics-Aware Trajectory Optimization via Convex Surrogates: We present a novel trajectory design for the jammer UAV that directly integrates a realistic rotary-wing propulsion energy model; By introducing the auxiliary variables for the speed and the induced power, we transform the non-convex propulsion function into a convex Second-Order Cone Program (SOCP). This is coupled with the first-order affine upper bounds on the eavesdropper rate as a function of the jammer’s position. This allows the algorithm to dynamically trade off a modest increase in the flight path for a significant gain in the jamming efficiency; minimizing the total energy consumption in a way that the naive trajectory planning cannot. Aerodynamics-Aware Trajectory Optimization via Convex Surrogates: We present a novel trajectory design for the jammer UAV that directly integrates a realistic rotary-wing propulsion energy model; By introducing the auxiliary variables for the speed and the induced power, we transform the non-convex propulsion function into a convex Second-Order Cone Program (SOCP). This is coupled with the first-order affine upper bounds on the eavesdropper rate as a function of the jammer’s position. This allows the algorithm to dynamically trade off a modest increase in the flight path for a significant gain in the jamming efficiency; minimizing the total energy consumption in a way that the naive trajectory planning cannot. Comprehensive Validation with 3GPP-Calibrated Simulations and Benchmarking: We provide an extensive performance evaluation using 3GPP UMa parameters and a realistic mining site geometry. Against four distinct baselines (Straight-Line Trajectory (SLT), Sequential Trajectory and Resource Optimization (STRO), Certainty-Equivalent Non-Robust Scheme (CENR), Static Hovering Jammer (SHJ)), our proposed scheme demonstrates:Superior EE: Achieving gains of up to 22.2% over STRO and 116.1% over SHJ.High Robustness: Maintaining only a 13.1% performance drop under severe eavesdropper uncertainty ( m), while a non-robust scheme collapses by 76.9%.Practical Viability: With under 2 seconds per iteration and fast convergence, the framework is suitable for near-real-time implementation, filling a critical gap between theoretical optimization and practical deployment in dynamic mining environments. Comprehensive Validation with 3GPP-Calibrated Simulations and Benchmarking: We provide an extensive performance evaluation using 3GPP UMa parameters and a realistic mining site geometry. Against four distinct baselines (Straight-Line Trajectory (SLT), Sequential Trajectory and Resource Optimization (STRO), Certainty-Equivalent Non-Robust Scheme (CENR), Static Hovering Jammer (SHJ)), our proposed scheme demonstrates:Superior EE: Achieving gains of up to 22.2% over STRO and 116.1% over SHJ.High Robustness: Maintaining only a 13.1% performance drop under severe eavesdropper uncertainty ( m), while a non-robust scheme collapses by 76.9%.Practical Viability: With under 2 seconds per iteration and fast convergence, the framework is suitable for near-real-time implementation, filling a critical gap between theoretical optimization and practical deployment in dynamic mining environments. Superior EE: Achieving gains of up to 22.2% over STRO and 116.1% over SHJ. Superior EE: Achieving gains of up to 22.2% over STRO and 116.1% over SHJ. High Robustness: Maintaining only a 13.1% performance drop under severe eavesdropper uncertainty ( m), while a non-robust scheme collapses by 76.9%. High Robustness: Maintaining only a 13.1% performance drop under severe eavesdropper uncertainty ( m), while a non-robust scheme collapses by 76.9%. Practical Viability: With under 2 seconds per iteration and fast convergence, the framework is suitable for near-real-time implementation, filling a critical gap between theoretical optimization and practical deployment in dynamic mining environments. Practical Viability: With under 2 seconds per iteration and fast convergence, the framework is suitable for near-real-time implementation, filling a critical gap between theoretical optimization and practical deployment in dynamic mining environments. The remainder of this paper is organized as follows: Section System Model and Problem Formulation introduces the system model and problem formulation; Section Proposed Solution details the proposed joint optimization solution; and Section Simulation Results analyzes the simulation results. Finally, Section Conclusion concludes the paper. For convenient reference, key mathematical notations and acronyms are summarized in Tables 2 and 3, respectively. Table 2Summary of key mathematical notations.SymbolDescriptionSymbolDescriptionSets and IndicesOptimization Blocks & ObjectivesU, uNumber of users, user indexRadio, FDMA, offloading blockL, ℓNumber of eavesdroppers, Eve indexLocal and edge CPU blockN, nNumber of time slots, slot indexJammer trajectory blockServer index (BS or AES)Lagrangian/Augmented objectiveΔTime slot durationηDinkelbach parameterDecision VariablesAuxiliary & Slack VariablesJammer 3D trajectory coordinateAuxiliary throughputOffloading indicator to server sRobust eavesdropper epigraphLocal CPU frequencySecrecy-fraction variableEdge CPU frequency for user uRatio auxiliary variableUser and jammer transmit powerV[n], y[n]Speed surrogate, induced power slackSystem ParametersChannel & Physical ModelsUser task size (bits)Eavesdropper data rateBandwidth allocationJammer-to-Eve channel gainκEffective switched capacitanceTrue and estimated Eve coordinatesRotary-wing propulsion constantsEavesdropper uncertainty radius Summary of key mathematical notations. Table 3List of acronyms.AcronymDefinitionAcronymDefinitionSystem & ArchitectureAESAerial Edge ServerMECMobile Edge ComputingBSBase StationPLSPhysical Layer SecurityCPUCentral Processing UnitUAVUnmanned Aerial VehicleEEEnergy EfficiencyUMaUrban Macro (3GPP Environment)Algorithms & OptimizationAOAlternating OptimizationMINLPMixed-Integer Non-Linear ProgramBCDBlock Coordinate DescentMMMajorization-MinimizationKKTKarush-Kuhn-TuckerQCPQuadratically Constrained ProgramLPLinear ProgramQCQPQuadratically Constrained Quadratic ProgramSCASuccessive Convex ApproximationSOC(P)Second-Order Cone (Program)Communications & Physical LayerA2GAir-to-GroundMIMOMultiple-Input Multiple-OutputCSIChannel State InformationNLoSNon-Line-of-SightFDMAFrequency Division Multiple AccessRFRadio FrequencyISACIntegrated Sensing and CommunicationSINRSignal-to-Interference-plus-Noise RatioLoSLine-of-SightIRSIntelligent Reflecting SurfaceSimulation Baselines (Benchmarks)CENRCertainty-Equivalent Non-RobustSLTStraight-Line TrajectorySHJStatic Hovering JammerSTROSequential Trajectory & Resource Opt. List of acronyms. System model and problem formulation We consider an offline setting where task parameters for all users and slots are known at the beginning of the horizon. Eavesdroppers are ground-based with zero altitude. The jammer is assumed to remain active throughout the slot. To ensure mathematical consistency with 3GPP standards, the small-scale fading is strictly state-conditioned (Rician for LoS and Rayleigh for Non-Line-of-Sight (NLoS)). Network architecture As illustrated in Fig. 1, the system comprises a ground BS and an AES that collectively serve a set of aerial users. The coordinate system defines the horizontal plane and respective altitudes. Let denote the set of users. The BS is located at horizontal coordinate with altitude , while the AES (e.g., a High-Altitude Platform Station (HAPS)) is positioned at with altitude . Each user is at horizontal coordinate and common altitude . Practically, the ground BS maps to the localized mine control center, while the AES functions as a high-altitude platform guaranteeing LoS coverage over the open pit. Proposed spatial architecture and PLS threat model. Multiple user UAVs offload smart mining tasks to a ground BS or an AES, while a dedicated jammer UAV transmits artificial noise along its trajectory to secure uplink transmissions against a worst-case ground eavesdropper. A protective jammer UAV flies at fixed altitude along a horizontal trajectory over the N slots of duration Δ. This jammer patrols the operational perimeter to secure proprietary data uplinks by blinding unauthorized ground receivers (eavesdroppers) attempting corporate espionage. Denote by with the jammer’s position in slot n. We define the slot-average speed as . The trajectory is constrained by initial and final waypoints and , maximum speed , and maximum acceleration to ensure physically realizable flight profiles: The system faces L passive eavesdroppers on the ground, . Their true horizontal positions are unknown; only estimates are available, subject to bounded uncertainty , where is the uncertainty radius (Practically, the estimated coordinates are acquired via UAV optical/thermal surveillance, active perimeter radar, or by detecting inadvertent RF leakage from the local oscillators of passive receivers. The mathematical bound m encapsulates the maximum spatial estimation error and resolution limits of these sensors in heavily obstructed, dust-prone mining environments). Channel and signal model The physical distances (in meters) , , , and are geometry-induced and defined as: We adopt the 3GPP TR 36.777 probabilistic channel model for an UMa environment. The probability of establishing a LoS connection for a generic A2G link is: where θ is the elevation angle in degrees. For the UMa scenario, we adopt standard parameters a = 4.32 and b = 0.11. The expected large-scale channel gain is: where and . The carrier frequency is GHz, the speed of light is c, and the state-dependent excess attenuation factors are dB and dB. Small-scale fading is strictly tied to the link state (TR 38.901). LoS links experience Rician fading, while NLoS links experience Rayleigh fading. Assuming the AES operates as a HAPS well above user UAV altitudes (), the Air-to-Air (A2A) link exhibits a dominant LoS probability (). The instantaneous channel power gains are denoted as , , , and . Additive white Gaussian noise powers are denoted by , , and . Communication rate model We adopt a Frequency Division Multiple Access (FDMA) protocol15 where the total system bandwidth W is dynamically divided among users. Let denote the bandwidth allocated to user u, subject to and . The SINRs are given by: The corresponding spectral efficiencies (in bits/s/Hz) follow the Shannon formula: Binary offloading indicators satisfy . The transmission time for a task of bits is: Computation model Tasks require CPU cycles. Local computing utilizes CPU frequency bounded by . Edge computing allocates CPU rates and from the BS and AES respectively, bounded by sum-capacities and . To ensure tasks are completed in real-time, the computation scheduling is governed by an intra-slot causality constraint, as illustrated in Fig. 2. Specifically, for edge offloading, the execution of these CPU cycles at the BS or AES can only commence after the uplink transmission phase has finished, with the total delay strictly bounded by the slot duration Δ. Intra-slot execution timing and causality constraints. The diagram illustrates the mutual exclusivity of Mode 0 (local computing) and Mode 1 (edge offloading) within a time slot Δ. For Mode 1, strict sequential causality ensures the uplink transmission completes before edge CPU execution begins, while the jammer UAV provides continuous secrecy protection. To ensure intra-slot task completion without queuing, local computing must satisfy . For edge processing, assuming strict causality (no overlap between transmission and computation), we enforce: Secrecy model Robust secrecy against the worst-case eavesdropper location is defined as: Remark 1 (Worst-case location) The inner maximization in (12) is attained on the boundary of the uncertainty ball due to monotonic distance-dependent path loss. In practice, the extremal point aligns with directions that increase the user–eavesdropper gain and/or decrease the jammer–eavesdropper gain; we evaluate candidate boundary directions within the SCA loop to obtain a safe upper bound for . To prevent optimization infeasibility under severe uncertainty, we introduce non-negative slack variables for the minimum secrecy Quality of Service (QoS): Secure bits processed are . Energy consumption model Communication & Jamming: Uplink energy is . Jamming Radio Frequency (RF) energy is . UAV Propulsion42: The propulsion energy per slot is , where is a standard rotary-wing propulsion power model (e.g., , with parameters defined in Table 4). The total jammer energy budget couples RF and propulsion: . Table 4Comprehensive system parameters.Parameter (Sym.)ValueParameter (Sym.)ValueNetwork & CommunicationComputation & MECSystem bandwidth (W)20 MHzMax CPU BS ()20 GHzTime slots (N)50Max CPU AES ()10 GHzSlot duration (Δ)1 sMax CPU User ()2 GHzCarrier freq. ()2 GHzSwitched cap. (κ)J/cycNoise power ()−110 dBmSecrecy penalty (λ) Radio & Channel ModelsUAV Kinematics42Max user power ()0.2 WMax UAV speed ()20 m/sMax jam power ()1 WMax UAV accel. ()5 m/sMin secrecy rate ()0.5 bps/HzJammer budget ()JExcess atten. (η)1, 20 dBProp. const ()79.8, 0.01, 88.63GPP UMa param. (a, b)4.32, 0.11 Comprehensive system parameters. Computation: Assuming multi-core isolation at the edge servers (independent cores per user with dedicated Dynamic Voltage and Frequency Scaling (DVFS)), the computation energy is: Problem formulation Before presenting the mathematical objective, we illustrate the logical relationship between the system parameters and the fractional EE goal in Fig. 3. This architecture demonstrates how task data is routed via offloading decisions, how energy components are aggregated, and how the PLS domain protects the data. Logical data and energy flow. This diagram illustrates the mapping of sensor data and offloading decisions to the system EE ratio, balancing aggregated energy costs against PLS-protected throughput. Let denote the decision variables. To account for secrecy outages via the slack variables, we formulate the robust EE maximization with a hefty penalty parameter λ> 0: Constraints (1)–(3) ensure the jammer UAV adheres to physically realizable flight paths bounded by maximum speed and acceleration, while meeting predefined initial and final waypoints. Constraint (11) enforces strict task causality, guaranteeing that edge computation only commences after the task data is fully transmitted. The robust secrecy requirements are governed by (13), which ensures a minimum secure rate during offloading, while the non-negative slack variables defined in (21) maintain feasibility under severe uncertainty. Constraint (16) enforces the FDMA bandwidth-sharing protocol, and (17) restricts the peak transmit powers for all users and the jammer. The jammer’s total energy budget, coupling both RF emissions and aerodynamic propulsion, is restricted by (18). Constraints (19) dictate that the task offloading decisions are binary and mutually exclusive. Finally, the computational capacities for the edge servers and local users are bounded by (20) and (21), respectively. Solution roadmap: We apply Dinkelbach’s transform to handle the fractional objective in (P1), and adopt AO across , , and . Non-convex rate and secrecy terms are linearized via first-order SCA around the current iterate, and binary offloading variables are time-sharing relaxed to [0, 1] with a final rounding step. Proposed solution In this section, we solve the joint problem defined in the System Model without restating it. We first remove the fractional structure via Dinkelbach’s transform, then employ a three-block BCD with conservative (inner) SCA surrogates. All symbols are those already introduced; new auxiliaries are explicitly declared when needed. Moreover, the detailed steps are illustrated in Fig. 4. Flowchart of the robust Dinkelbach-BCD-SCA algorithm. The architecture decouples the non-convex MINLP into three tractable subproblems (AO, Quadratically Constrained Quadratic Program (QCQP), and SOCP) within an Majorization-Minimization (MM) framework, followed by a final feasibility repair. Problem decomposition via Dinkelbach–BCD Dinkelbach transform. Let η>0 be a parameter. The parametric objective is subject to all constraints of the original problem. For a given η, we solve Define . Dinkelbach’s updates are , with termination when . Theorem 1 (Dinkelbach equivalence) Let be a stationary solution of the original fractional problem. Then satisfies . Conversely, if , any maximizer of is stationary for the original problem. Proof Follows from standard fractional programming arguments: F is continuous, strictly decreasing, and iff equals the optimal ratio; details omitted for brevity. □ Block partition (BCD). We partition variables into three blocks: Per Dinkelbach iteration, BCD cyclically optimizes one block while freezing the other two at their latest values. Each block is convexified via provably conservative surrogates, yielding monotone ascent of . Subproblem : radio, FDMA, offloading, and slacks Subproblem definition For fixed , we optimize augmented with the auxiliary throughput variables , robust eavesdropper epigraphs , secrecy-fraction variables for , and ratio auxiliaries . The exact parts of the constraints retained from the System Model are not repeated here. The block objective collects all terms in that depend on and the auxiliaries. Non-convexity analysis The following terms prevent direct convex optimization: (i) the throughput product with non-linear ; (ii) the secrecy max via ; (iii) the secure-bit ratio in ; (iv) bilinear energy appears convex but is coupled with the non-convex throughput constraint. Convex transformation (a) Robust eavesdropper epigraph. Introduce for every (u, n) the variable and enforce This is an exact reformulation of via epigraph. For tractability, we linearize each at : where, with , , Substituting (25) into (24) yields affine secrecy constraints. (b) Throughput epigraph and micro-AO. Introduce and impose For fixed , the mapping is a positive scaling of the concave , hence the hypograph is convex. For fixed , is constant and (26) is linear in . Therefore, we solve Block with a two-sweep alternating update: (c) Secrecy ratio lower bound. Introduce and and enforce Lemma 1 If , then , thus is a lower bound on the secure offloaded bits. Proof From , . □ The epigraph in (27) is convex since is concave and x↦ 1/x is convex and nonincreasing on (composition rule). (d) Communication energy convexity. By the perspective property, (p,t)↦ p/t is convex on ; hence is convex. Final convex subproblem At a given sweep (A or B), the convex program is To resolve the bi-convexity of the throughput capacity constraint, we cannot solve the radio resource subproblem in a single shot. Instead, we embed a micro-AO procedure within the block’s SCA framework. Specifically, we alternate between optimizing radio resources and offloading indicators (Sweep A) while freezing the bandwidth, and subsequently optimizing the bandwidth allocation (Sweep B) while freezing the radio resources. Because each sweep solves a strictly convex inner approximation, the local objective monotonically non-decreases. This iterative two-sweep procedure is summarized in Algorithm 1. Block- (Two-Sweep AO + SCA) Subproblem : local and edge CPU allocation For fixed , the optimization of the local and edge CPU frequencies, denoted by , decouples from the radio and trajectory variables. Because the offloading indicators and the throughput variables are provided as fixed parameters from Block 1, the objective function reduces to minimizing the computation energy. This formulation is strictly convex quadratic with respect to and . Furthermore, the causality and task completion constraints depend on the edge CPU rates via the inverse function , which is strictly convex for . Consequently, without requiring any successive convex approximations or Taylor expansions, the subproblem is intrinsically a convex QCQP given by: Because is strictly convex, it can be solved to global optimality in a single pass using standard interior-point methods, as summarized in Algorithm 2. Block- (Convex QCQP) Subproblem : Jammer trajectory optimization With the resource and computation blocks fixed, the optimization of the jammer trajectory is fully decoupled from the user-side energy metrics. The objective strictly minimizes the jammer’s propulsion energy, subject to the UAV ’s kinematics and robust secrecy requirements. Non-convexity analysis and SCA for secrecy constraints The first non-convexity in this subproblem arises from the eavesdropper rate , which depends on the dynamic jammer-to-eavesdropper channel gain . Assuming a path-loss exponent , the channel gain itself is convex with respect to the position vector ; however, its position in the denominator of the SINR within the logarithmic rate function renders the secrecy constraint non-convex. To guarantee a conservative inner approximation, we construct a globally valid affine upper bound for the eavesdropper rate using a first-order Taylor expansion around the local trajectory iterate . Applying the multidimensional chain rule, the gradient with respect to the jammer position is: where the constituent directional derivatives evaluated at the local point are given by: Consequently, the non-convex eavesdropper rate is strictly upper-bounded by the affine surrogate: By enforcing the epigraph condition for all eavesdroppers , the feasible region of the worst-case secrecy constraint is transformed into a convex polyhedron. SCA for aerodynamic propulsion energy The second severe non-convexity stems from the rotary-wing propulsion power model, defined as , where . While is convex, the induced power term is non-convex with respect to the trajectory vector . To resolve this, we introduce two auxiliary variables: a bounded speed surrogate V[n] and an induced power slack variable y[n]. We reformulate the propulsion power as a convex function: subject to the condition y[n] \ge 1/V[n], which can be exactly represented as a convex Second-Order Cone (SOC) constraint: To strictly couple V[n] with the physical velocity, we require . Because this constraint defines a non-convex feasible set, we apply a first-order Taylor expansion on the right-hand side around the local trajectory velocity , yielding the affine lower bound: Final convex subproblem and algorithm Integrating the secrecy surrogates and the aerodynamic SOC constraints yields the following strictly convex Quadratically Constrained Program (QCP) for the trajectory: Because relies on local first-order approximations, we must iteratively update the trajectory point within an SCA loop until convergence, as detailed in Algorithm 3. Block- (Trajectory SCA with Aerodynamics) Master algorithm, complexity, and convergence analysis Let and . The Dinkelbach framework seeks the root of the auxiliary function , where encapsulates the continuous feasible space defined by the constraints in . The iterative zero-finding procedure and the embedded BCD are formalized in Algorithm 4. Dinkelbach–BCD–SCA Master Procedure Computational complexity The overall complexity is governed by the worst-case arithmetic operations of the interior-point method used to solve the convex subproblems. For a generic convex program with variables and constraints, the iteration complexity scales as . Block : Sweep A contains variables and epigraph constraints, yielding a complexity of . Sweep B is an Linear Program (LP) scaling as . Block : Sweep A contains variables and epigraph constraints, yielding a complexity of . Sweep B is an Linear Program (LP) scaling as . Block : The QCQP involves variables and constraints, yielding . Block : The QCQP involves variables and constraints, yielding . Block : The SOCP involves positional variables and constraints, yielding . Block : The SOCP involves positional variables and constraints, yielding . Let and denote the number of outer Dinkelbach and inner BCD iterations, respectively. The asymptotic total computational complexity is: Crucially, this worst-case complexity is strictly polynomial rather than exponential. By relaxing the binary offloading indicators to continuous variables and resolving the subproblems via convex interior-point methods, the framework explicitly avoids the combinatorial explosion (e.g., ) inherent to exact mixed-integer solvers. Coupled with execution at the high-capacity ground BS and the rapid stabilization demonstrated within the first few iterations, this polynomial scaling strictly guarantees computational tractability for near-real-time implementations in dynamic mining environments. Convergence analysis We establish convergence by invoking the strict properties of the MM framework. Let represent the non-convex objective of any respective block and be its surrogate evaluated at the local iterate . Proposition 2 (SCA Surrogate Tightness) The affine and quadratic surrogates constructed in Section Proposed Solution satisfy the following strict MM conditions : Proof Eq. (39) holds globally due to the concavity of the Shannon rate functions and the convexity of the 1/x transform applied to the fractional terms. Equations (40) and (41) hold trivially as the surrogates are derived via exact first-order Taylor expansions evaluated precisely at . □ Theorem 3 (Stationarity and Sequence Convergence) The sequence generated by Algorithm 4 monotonically converges to a Karush-Kuhn-Tucker (KKT) stationary point of the relaxed formulation of . Proof By Proposition 2, the inner BCD update guarantees . Consequently, the sequence of objectives is monotonically non-decreasing. Given that is a compact set and the energy denominator is strictly bounded away from zero, the sequence is bounded above. By the Monotone Convergence Theorem, the inner loop converges. Furthermore, because is strictly concave and is strictly convex under the SCA transformations, the parametric Dinkelbach update yields a superlinear convergence rate to such that . By the continuous differentiability of the constraint set, the limit point satisfies the KKT conditions of the relaxed problem. Crucially, because the MM surrogates serve as strict conservative lower bounds, the update steps are mathematically insulated against the trajectory oscillations and divergence commonly associated with highly coupled optimization problems. This decoupled iterative architecture ensures structural stability and rapid convergence–a paradigm consistent with recent multi-variable UAV optimization frameworks43 guaranteeing steady progression to a stationary point. □ Remark 2 (Impact of Relaxation and Feasibility Repair) While Theorem 3 establishes the exact convergence of the continuous relaxation, the original problem is a MINLP with binary offloading indicators . Relaxing these to guarantees polynomial tractability. Let denote the stationary point of this relaxed feasible space, where represents the continuous variables. Mapping back to the discrete domain via magnitude thresholding introduces a rounding perturbation , which risks violating tightly coupled physical constraints (such as intra-slot causality and peak CPU limits). To strictly guarantee physical feasibility, Algorithm 4 executes a final repair phase. By fixing the rounded binary indicators and executing a final projection pass of the strictly convex resource blocks ( and ), the continuous variables are deterministically adjusted to absorb the perturbation . Because Blocks and are strictly convex given a fixed , this step acts as a secure projection onto the exact binary feasible set , restoring constraint strictness without incurring exponential branch-and-bound complexity. Simulation results This section evaluates the performance of the proposed robust joint EE maximization framework. We consider a 3D Urban Macro (UMa) environment spanning a 1000 × 1000 m area. Unless otherwise specified, the default system and algorithmic parameters are summarized in Table 4. The network comprises U = 4 legitimate aerial users randomly distributed within a 200 m radius of the center at a common altitude m. The ground BS is centrally located at m with an antenna height m, while the AES hovers directly above at m to ensure dominant Los Air-to-Air links. The system faces L = 2 passive ground eavesdroppers with estimated coordinates m and m. To test the robust formulation, their true locations are randomly perturbed within an uncertainty radius m. The protective jammer UAV operates at m, navigating from an initial waypoint m to a final destination m. User task sizes are uniformly distributed in [0.5, 2] Mbits with a computation intensity of 1000 CPU cycles/bit. The framework is implemented in MATLAB utilizing the MOSEK solver via CVX. To rigorously demonstrate the advantages of our proposed Dinkelbach-BCD solution, we first define several comparative baseline schemes. Benchmark schemes To rigorously evaluate the performance of the proposed Dinkelbach-BCD based robust joint optimization scheme, we compare it against four representative baseline schemes. These benchmarks are designed to support ablation studies on trajectory design, algorithmic coupling, robust worst-case secrecy formulation, and jammer mobility, consistent with current trends in secure UAV-assisted MEC9,14. Straight-Line Trajectory (SLT): This scheme evaluates the impact of dynamic trajectory optimization5,19. The jammer UAV is forced to fly along a direct path from the initial waypoint to the final destination at a constant speed. Mathematically, the trajectory is predetermined as for all . The remaining radio, offloading, and computation variables are optimized using Blocks and of the proposed algorithm. Straight-Line Trajectory (SLT): This scheme evaluates the impact of dynamic trajectory optimization5,19. The jammer UAV is forced to fly along a direct path from the initial waypoint to the final destination at a constant speed. Mathematically, the trajectory is predetermined as for all . The remaining radio, offloading, and computation variables are optimized using Blocks and of the proposed algorithm. Sequential Trajectory and Resource Optimization (STRO): This scheme demonstrates the necessity of the proposed iterative BCD framework. Instead of cyclically updating the blocks until convergence, the problem is decoupled sequentially. First, the jammer trajectory is optimized (via Block ) assuming uniform power allocation. Then, with fixed, the radio and computation resources are optimized in a single pass of Blocks and . Sequential Trajectory and Resource Optimization (STRO): This scheme demonstrates the necessity of the proposed iterative BCD framework. Instead of cyclically updating the blocks until convergence, the problem is decoupled sequentially. First, the jammer trajectory is optimized (via Block ) assuming uniform power allocation. Then, with fixed, the radio and computation resources are optimized in a single pass of Blocks and . Certainty-Equivalent Non-Robust Scheme (CENR): To validate the proposed worst-case secrecy formulation, this baseline ignores the eavesdropper location uncertainty10. It assumes the estimated locations are perfectly accurate by setting the uncertainty radii . Consequently, the robust inner maximization in the PLS rate is removed. During simulation testing, the true locations are subjected to random perturbations to measure the resulting secrecy outages. Certainty-Equivalent Non-Robust Scheme (CENR): To validate the proposed worst-case secrecy formulation, this baseline ignores the eavesdropper location uncertainty10. It assumes the estimated locations are perfectly accurate by setting the uncertainty radii . Consequently, the robust inner maximization in the PLS rate is removed. During simulation testing, the true locations are subjected to random perturbations to measure the resulting secrecy outages. Static Hovering Jammer (SHJ): This scheme tests the physical architecture by turning off the mobility of the jammer16. The UAV jammer is deployed at a fixed, central location for the entire horizon, i.e., , where is the geometric center of the user distribution. The propulsion energy strictly becomes , and the remaining variables are optimized jointly. Static Hovering Jammer (SHJ): This scheme tests the physical architecture by turning off the mobility of the jammer16. The UAV jammer is deployed at a fixed, central location for the entire horizon, i.e., , where is the geometric center of the user distribution. The propulsion energy strictly becomes , and the remaining variables are optimized jointly. Numerical discussion This section evaluates the performance of the proposed robust joint optimization framework through the numerical simulations. The default parameters, unless otherwise stated, follow those listed in the Table 4. Trajectory analysis Figure 5 visualizes the 3D network geometry and the spatial behavior of the jammer UAV trajectory over the operational horizon. The Straight-Line Trajectory (SLT) benchmark follows a direct path between the waypoints; covering a geometric distance of approximately 1131 m. In contrast, the Proposed scheme dynamically maneuvers the UAV to proactively intercept the eavesdropper locations while maintaining a spatial buffer from the legitimate user cluster. This optimized routing extends the total flight path to approximately 1250 m; a 10.5% increase in the total distance. This deviation reveals a fundamental physical trade-off; the system deliberately expends the additional aerodynamic propulsion energy to fly the longer route because the closer proximity to the eavesdroppers provides an exponential increase in the jamming channel gain ; thereby securing the offloading links. 3D jammer UAV trajectory and system geometry. System scalability and robustness Figure 6 illustrates the system EE scalability as the number of the legitimate users (U) increases from 2 to 10. The Proposed Joint Scheme consistently dominates; peaking at 60.5 kbits/J at U=10. Specifically, the Proposed scheme outperforms the STRO benchmark (49.5 kbits/J) by 22.2% and the SHJ baseline (28.0 kbits/J) by an overwhelming 116.1%. This superiority arises because the iterative BCD framework synchronizes the jammer’s kinematic positioning with the users’ heterogeneous bandwidth and the computational requirements. System EE versus number of users U. Figure 7 evaluates the resilience against the location uncertainty . At the m, the Proposed scheme maintains 45.2 kbits/J; representing a minor 13.1% degradation from the perfect-information case. Conversely; the non-robust CENR benchmark plummets to 12.0 kbits/J, a catastrophic 76.9% drop. This occurs because the CENR optimizes for the estimated center coordinates; as the true eavesdroppers fall outside the narrow jamming beam, the system suffers the severe secrecy outages. The Proposed scheme avoids this by expanding the jamming coverage to blanket the entire uncertainty ball. Robust EE versus eavesdropper uncertainty . Secrecy and energy trade-offs Figure 8 demonstrates the security trade-off by plotting the EE against the minimum secrecy rate . As the quality of service demands become stricter; the EE drops because the jammer must either fly closer to the eavesdroppers or increase the RF power w[n]. When the increases from 0.1 to 0.9 bits/s/Hz; the Proposed scheme’s EE drops by 42.5%, whereas the SHJ architecture collapses by 88.5%. At the most stringent requirement (); the proposed optimization operates 117.2% more efficiently than the SLT baseline, confirming that the dynamic mobility is required to fulfill the high-security offloading. Energy efficiency versus minimum secrecy rate . Figure 9 dissects the total energy consumption components. The aerodynamic propulsion dominates the budget; accounting for 4800 J (86.8% of the total) in the Proposed scheme. Comparatively, while the SLT baseline saves 6.3% in the propulsion energy; its sub-optimal positioning forces its RF jamming energy to skyrocket by 730.9% to 2742 J. This validates that optimizing the drone mobility specifically minimizes the overall energy by reducing the need for the high-power; long-range RF radiation. Breakdown of total energy consumption components. Architecture and complexity Figure 10 presents an analysis of the dual-edge computing. Under the light workloads; the AEO benchmark (65.0 kbits/J) outperforms the GEO benchmark (55.2 kbits/J) by 17.8% due to the superior LoS channels. However, under the extreme loads (2.0 Mbits), the AES CPU becomes overwhelmed; causing the AEO to crash to 18.2 kbits/J. In this state, the GEO surpasses the AEO by 40.1% due to the BS’s larger computational capacity. The proposed dual-edge architecture achieves 31.5 kbits/J under the extreme loads; outperforming the GEO and the AEO by 23.5% and 73.1%, respectively, by intelligently splitting the task load. Energy efficiency across various task workload scenarios. Figure 11 evaluates the computational complexity. The Full Proposed Joint Optimization requires an average of 1.95 s per iteration. Fixing the trajectory () reduces the time to 0.98 s (a 49.7% reduction); while freezing the radio () or the CPU () only saves 33.3% and 25.6%, respectively. This indicates that resolving the SOCPs for the propulsion and the spatial Taylor approximations is the primary bottleneck. Nevertheless; a runtime under 2 s validates the framework’s feasibility for the near-real-time implementation. Average execution time per BCD iteration. Probabilistic and convergence analysis Figure 12 provides the probabilistic verification via the Monte Carlo simulations. Given the ; the Proposed scheme achieves a 0% outage probability, proving the absolute protection against the worst-case boundary. In contrast, the CENR benchmark violates the secrecy constraints in 65% of the realizations at the m. This proves that the robust epigraphs are essential for maintaining the secrecy feasibility in the presence of the spatial estimation errors. Cumulative distribution function (CDF) of the secrecy rate. Finally; Fig. 13 shows the convergence trends for the varying user populations. A strict monotonic increase confirms the theoretical guarantees of the MM framework. The algorithm exhibits the rapid convergence; for the U=6, the efficiency stabilizes by the seventh iteration, capturing a 113.7% relative improvement. Crucially, the system absorbs 85% of its total gain within the first three iterations; demonstrating that the framework is computationally stable and scalable under the dense network conditions. Convergence of the proposed Dinkelbach-BCD framework. To evaluate practical viability, Fig. 14 analyzes system sensitivity to maximum UAV speed () and altitude (). A clear non-linear trade-off exists for speed: energy efficiency peaks near 15 m/s, balancing the benefit of reaching optimal defensive waypoints quickly against the severe propulsion penalty of profile drag, which scales cubically at higher speeds. Furthermore, increasing altitude from 50 m to 150 m monotonically degrades efficiency; the severe free-space path loss introduced by greater absolute distances overpowers any line-of-sight improvements, forcing the radio circuitry to consume substantially more power to satisfy the strict secrecy constraints. System energy efficiency versus maximum UAV speed and altitude . Conclusion This work presented a realism-aware framework for secure aerial-edge computing in intelligent mining that jointly optimizes radio resources, computation scheduling, and a protective jammer UAV trajectory under 3GPP-calibrated A2G channels, bounded eavesdropper uncertainty, and rotary-wing propulsion physics. The robust secrecy formulation combines epigraphed worst-case eavesdropper rates with non-negative slack variables to preserve feasibility, while strict slot-level causality guarantees transmission-then-computation ordering. To solve the resulting fractional and highly coupled problem efficiently, we integrated Dinkelbach’s transform with a three-block BCD and conservative SCA surrogates, including a micro alternating-optimization within the FDMA radio block and an SOC representation for induced power in the aerodynamics model. Comprehensive evaluations in an Urban Macro setting demonstrate that the proposed method consistently outperforms Straight-Line (SLT), Sequential (STRO), Non-Robust (CENR), and Static Hover (SHJ) baselines in system EE, maintains robustness against eavesdropper location uncertainty with modest degradation at m, and converges within practical computational budgets (sub-2 s average per iteration with rapid monotonic ascent), thereby supporting near-real-time deployment. Analyses further reveal that propulsion is the dominant energy component and show that geometry-aware jamming can suppress RF energy by positioning rather than solely by power. Future directions include: (i) online and learning-augmented variants for time-varying traffic and channels; (ii) multi-UAV cooperative jamming and distributed dual-edge orchestration; (iii) richer uncertainty sets (e.g., distributionally robust secrecy and partial Channel State Information (CSI)) and stochastic task arrivals; and (iv) integration with advanced physical-layer enablers (e.g., massive MIMO, IRS, or ISAC) and safety-critical field trials in operational mines. These avenues can further close the gap between theoretical guarantees and large-scale, secure, and energy-efficient industrial deployments. Furthermore, while this paper establishes a theoretically rigorous mathematical optimization framework, exploring data-driven and learning-based techniques such as DRL to achieve faster, real-time trajectory and resource allocation in highly dynamic environments remains a highly promising direction for future work. Acknowledgment The authors extend their appreciation to Northern Border University, Saudi Arabia, for supporting this work through project number (NBU-CRP-2026-2903). Author contributions A.A.D. and A.A.A. conceptualized the study and developed the core methodology. K.A. and A.A.Z. performed the formal mathematical analysis and implemented the optimization algorithms in software. L.A.D. conducted the simulation runs and prepared the figures. K.N.R.A. contributed to the robust secrecy modeling and validation. A.A.D. wrote the main manuscript text. A.A.A., K.A., and K.N.R.A. extensively reviewed and edited the manuscript. A.A.A. supervised the project and acquired funding. All authors reviewed and approved the final manuscript. Funding The authors extend their appreciation to the Deanship of Scientific Research at Northern Border University, Arar, KSA for funding this research work through the project number “NBU-CRP-2026-2903”. Data availability The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request. The MATLAB implementation of the proposed framework used in this study is available from the corresponding author upon reasonable request. Code availability The MATLAB implementation of the proposed framework used in this study is available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Consent for publication All authors have read and approved the final manuscript for publication. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request. The MATLAB implementation of the proposed framework used in this study is available from the corresponding author upon reasonable request. The MATLAB implementation of the proposed framework used in this study is available from the corresponding author upon reasonable request.