Low-altitude UAV Friendly-Jamming for Satellite-Maritime Communications via Generative AI-enabled Deep Reinforcement Learning

A LEO satellite-maritime communication system assisted by low-altitude UAV.

AI-Powered Drones Secure Maritime Satellite Communications Against Cyber Eavesdropping

Revolutionary system combines generative AI with autonomous UAVs to protect vulnerable ocean data transmissions

Researchers have developed a novel AI-powered system using low-altitude UAV friendly-jamming to secure LEO satellite-maritime communications against eavesdropping threats, achieving near-optimal security performance while minimizing energy consumption through a transformer-enhanced deep reinforcement learning approach.

As global maritime commerce increasingly relies on satellite communications for critical data transmission, a team of international researchers has unveiled a breakthrough security solution that deploys AI-controlled drones to jam potential eavesdroppers while preserving legitimate communications.

The research, led by scientists from Jilin University in China and Nanyang Technological University in Singapore, addresses a pressing vulnerability in Low Earth Orbit (LEO) satellite-maritime networks: their open channels and extensive coverage make them susceptible to unauthorized interception by malicious actors at sea.

The Security Challenge

LEO satellites, orbiting between 500 and 2,000 kilometers above Earth's surface, have become essential infrastructure for maritime communications due to their low latency and high data transmission efficiency. However, their broad coverage creates significant security risks. Traditional cryptography-based security methods require frequent data encoding and decoding, posing challenges in energy-limited maritime environments when transmitting large data volumes.

"The extensive coverage of LEO satellites, combined with the openness of channels, can cause the communication process to suffer from security risks," the research team explains in their paper published on arXiv. Physical layer security (PLS) offers a more dynamic alternative, but implementing it for fast-moving LEO satellites places high demands on hardware computational resources.

The UAV Solution

The researchers' solution employs low-altitude unmanned aerial vehicles (UAVs) as mobile jamming platforms. These drones, equipped with single omni-directional antennas and optical cameras, detect eavesdropping vessels and transmit targeted jamming signals to interfere with unauthorized signal reception while minimizing interference to legitimate users.

The system operates in a discrete-time framework where an LEO satellite transmits data signals to a legitimate vessel (Alice) over a communication link, while a potential eavesdropper (Eve) attempts to intercept the transmission. The UAV continuously adjusts its 3D position and transmission power to optimize jamming effectiveness.

The challenge lies in balancing conflicting objectives: maximizing the secrecy rate of communications while minimizing the UAV's energy consumption—particularly critical in maritime environments where recharging opportunities are limited.

---

SIDEBAR: Research Artifacts for Independent Verification

To enable independent verification and reproducibility, the research team has provided comprehensive technical specifications and methodological details:

Simulation Parameters

The complete parameter set is documented in Table 2 of the manuscript, including:

• Orbital parameters: Inclination angle (80°), right ascension of ascending node (70°), orbital altitude (900 km)
• Communication parameters: Satellite transmit power (49.03 dBm), antenna gains (satellite: 52 dBi, UAV: 8 dBi)
• Physical constraints: UAV altitude range (50-70m), maximum allowable interference power (-74 dBm), total communication energy budget (500 J)
• Channel modeling: Rician factor (31.3), path loss parameters for S2V link (CS=46.4, WS=2) and U2V link (CU=116.7, WU=1.5)

Algorithm Implementation Details

The TransSAC algorithm specifications include:

• Network architecture: Two hidden layers with ReLU activation, batch size of 128
• Hyperparameters: Learning rate (0.003), discount factor (0.9), soft update rate (0.5), transformer attention heads (8)
• Training protocol: 1×10⁶ total iterations with evaluation every 80 iterations
• Optimizer: Standard Adam optimizer with specified learning rates

Mathematical Models

The paper provides explicit equations for:

• LEO satellite orbit dynamics (Equation 1): 3D Cartesian coordinates based on Keplerian orbital elements
• Vessel movement (Equations 2-3): Six degree-of-freedom model with force vectors
• Channel characteristics (Equations 4-7): Composite Rician fading channels with path loss models
• Secrecy rate calculation (Equations 8-10): Information-theoretic security metrics
• UAV energy consumption (Equations 11-12): Propulsion power models incorporating horizontal and vertical flight dynamics

Baseline Comparisons

Comparative algorithm parameters are fully specified (Table 3):

• DDPG, TD3, PPO, and SAC configurations with matching batch sizes, learning rates, and discount factors
• Theoretical optimal secrecy rate calculation methodology for idealized scenarios

Performance Metrics

The study reports multiple quantifiable outcomes:

• Average secrecy rate approaching theoretical optimal (~8 bps/Hz)
• UAV energy consumption (~190-270 J over 40-second episodes)
• Convergence performance across 10,000 training iterations
• Constraint sensitivity analysis for power limits (18-26 dBm) and interference thresholds (-92 to -68 dBm)

Computational Environment

Hardware and software specifications:

• Server: AMD EPYC 7642 48-Core CPU, NVIDIA GeForce RTX 3090 GPU, 128 GB RAM
• Software: Python 3.8, Visual Studio Code 1.91
• Additional validation: Raspberry Pi implementation confirming real-time feasibility

Reproducibility Considerations

The researchers note that simulations account for:

• Vessel trajectory variability through stochastic environmental influences
• Dynamic random seed strategies across training episodes
• Variable LEO satellite initial positions to enhance state space exploration

Code and Data Availability

While the manuscript does not explicitly state a public code repository, the level of technical detail provided—including specific equations, network architectures, and parameter values—would allow expert practitioners to implement the system independently. The paper is available as an open-access preprint on arXiv (arXiv:2501.15468v2).

---

Generative AI Takes the Helm

To solve this complex optimization problem, the team developed the TransSAC (Transformer-enhanced Soft Actor-Critic) algorithm, a generative AI-enabled deep reinforcement learning approach. The algorithm incorporates a transformer architecture—the same technology underlying large language models like GPT—to capture temporal dependencies in the dynamic maritime environment.

The transformer component processes sequences of states and actions using self-attention mechanisms, enabling the system to understand how current decisions affect future outcomes across the entire communication session. This proves essential given that LEO satellites complete orbits every 90-120 minutes, vessels follow predetermined trajectories, and UAV flight paths must satisfy physical continuity constraints.

"The transformer-enhanced learning strategy can effectively address the challenge of strong temporal correlation through enhanced state and action representations while handling large-scale state and action spaces through parallelized processing," the researchers note.

Additionally, the system employs a multi-armed bandit (MAB) optimization scheme that dynamically explores different weight configurations for balancing the conflicting objectives, rather than relying on preset values that might prove suboptimal.

Impressive Performance Results

Simulation results demonstrate that the TransSAC algorithm significantly outperforms conventional deep reinforcement learning methods including DDPG (Deep Deterministic Policy Gradient), TD3 (Twin Delayed Deep Deterministic Policy Gradient), PPO (Proximal Policy Optimization), and standard SAC.

Most notably, the approach achieves a secrecy rate approaching the theoretical optimal value—a near-perfect result considering the multi-objective optimization constraints. The system also identified optimal constraint values for maximum UAV transmission power (20 dBm) and maximum allowable interference power (-74 dBm) that balance security performance with energy efficiency.

Compared to non-UAV approaches where LEO satellites transmit without jamming assistance, the UAV-assisted system consistently maintains superior secrecy rates, ensuring reliable communications even in the presence of sophisticated eavesdropping attempts.

Real-World Implications

The research has significant implications for maritime security across multiple sectors. Commercial shipping companies transmitting sensitive cargo and routing information, naval operations requiring secure tactical communications, offshore oil and gas facilities, and maritime research vessels collecting proprietary data could all benefit from enhanced communication security.

The system's energy efficiency proves particularly valuable given the challenges of recharging UAVs in maritime environments. By minimizing unnecessary position adjustments and optimizing transmission power, the approach extends operational duration while maintaining security performance.

Future Directions

The research team acknowledges that their current work assumes perfect knowledge of eavesdropper positions. Future research will explore scenarios with imperfect position information, leveraging deep reinforcement learning frameworks to predict illegitimate user locations.

The approach can also extend to multiple UAV scenarios, where coordinated drone swarms provide enhanced security performance and expanded maritime coverage capabilities—though this increases system complexity and coordination requirements.

Broader Context

This research exemplifies the growing convergence of artificial intelligence, autonomous systems, and cybersecurity. As satellite-based communications expand to support everything from autonomous shipping to ocean monitoring networks, securing these vulnerable transmission channels becomes increasingly critical.

The work also demonstrates how generative AI techniques originally developed for natural language processing can solve complex real-world optimization problems in wireless communications and network security.

With maritime trade accounting for over 80% of global commerce by volume, according to the United Nations Conference on Trade and Development, ensuring secure communications infrastructure represents both an economic and national security imperative.

---

Sources

1. Huang, J., Wang, A., Sun, G., Li, J., Wang, J., Niyato, D., & Leung, V. C. M. (2025). "Low-altitude UAV Friendly-Jamming for Satellite-Maritime Communications via Generative AI-enabled Deep Reinforcement Learning." arXiv preprint arXiv:2501.15468v2. https://arxiv.org/abs/2501.15468

2. Huang, J., Wang, A., Sun, G., Li, J., Wang, J., Du, H., & Niyato, D. (2025). "Dual AAV cluster-assisted maritime physical-layer secure communications via collaborative beamforming." IEEE Internet of Things Journal, 12(9), 12589-12607.

3. Alqurashi, F. S., Trichili, A., Saeed, N., Ooi, B. S., & Alouini, M. (2023). "Maritime communications: A survey on enabling technologies, opportunities, and challenges." IEEE Internet of Things Journal, 10(4), 3525-3547.

4. Zhou, D., Sheng, M., Li, J., & Han, Z. (2023). "Aerospace integrated networks innovation for empowering 6G: A survey and future challenges." IEEE Communications Surveys & Tutorials, 25(2), 975-1019.

5. Pokhrel, S. R., & Choi, J. (2024). "Data-driven satellite communication and control for future IoT: Principles and opportunities." IEEE Transactions on Aerospace and Electronic Systems, 60(3), 3307-3318.

[2501.15468] Low-altitude UAV Friendly-Jamming for Satellite-Maritime Communications via Generative AI-enabled Deep Reinforcement Learning