Abstract
As the Low-altitude Economy (LAE) expands, securing long-range drone-to-drone communication becomes a critical priority. This paper introduces GenSI, a Generative Swarm Intelligence framework designed to optimize Distributed Collaborative Beamforming (DCB) for Aerial Virtual Antenna Arrays (AVAAs). By integrating a Conditional Variational Autoencoder (CVAE) for "warm start" initialization with an enhanced Multi-Objective Ant Lion Optimizer (MOALO), GenSI achieves high-security beamforming robust against colluding eavesdroppers. Experimental results demonstrate a 58.7% reduction in execution time and superior secrecy performance compared to traditional methods, making it viable for real-time deployment on resource-constrained hardware.
Introduction
The expansion of the Low-altitude Economy (LAE) introduces critical security challenges for Autonomous Aerial Vehicles (AAVs) engaged in delivery and infrastructure inspection. While traditional upper-layer encryption is often too computationally heavy for drone hardware, Physical Layer Security (PLS) offers a lightweight alternative. However, PLS must contend with sophisticated threats like eavesdropper collusion, where multiple attackers coordinate their signal detection using techniques like Maximum Ratio Combining (MRC) to intercept data.
System Model and Problem Formulation
The proposed system utilizes two AAV swarms acting as Aerial Virtual Antenna Arrays (AVAAs) to establish a secure long-range link. The optimization problem is formulated as a multi-objective challenge, aiming to balance three conflicting goals:
- Maximizing Secrecy Capacity: Protecting data from known eavesdroppers.
- Minimizing Maximum Sidelobe Level (SLL): Reducing stray signal leakage to avoid detection by unknown eavesdroppers.
- Minimizing Energy Consumption: Reducing the movement required for AAVs to reach optimal "virtual antenna" positions.
The GenSI Framework
Solving this multi-objective problem is NP-hard. The GenSI framework addresses this complexity through a two-phase process:
1. CVAE-based Initialization (The "Warm Start")
Instead of starting the optimization from scratch, a Conditional Variational Autoencoder (CVAE) is trained on "expert" solutions from previous environments. When faced with a new scenario, the CVAE generates high-quality initial solutions, bypassing the time-consuming random exploration phase typical of standard heuristic algorithms.
2. Enhanced MOALO Evolution
The evolution phase uses an improved Multi-Objective Ant Lion Optimizer (MOALO). Key enhancements include a sorting-based population evolution to prevent biased solutions and a new integer update method to handle discrete variables like receiver selection.
Key Results and Insights
The framework was evaluated using synthetic datasets and a hardware testbed (Raspberry Pi 4B). Key findings include:
- Computational Efficiency: The CVAE mechanism resulted in a 58.7% reduction in execution time compared to a "cold start" approach.
- Security Performance: GenSI consistently outperformed baselines, reaching a secrecy capacity of ~2.12 Mbps while maintaining lower energy costs (~7.68 x 10⁴ J).
- Real-world Viability: The framework proved capable of running on a Raspberry Pi 4B in roughly 27 seconds, making it feasible for near-real-time updates in the field.
Conclusion and Future Work
GenSI provides a robust and efficient solution for secure aerial communications in the LAE. While the current method is best suited for moderately sized swarms, future work will explore scaling to larger swarms, addressing high-speed dynamic environments, and integrating functions like Integrated Sensing and Communication (ISAC).
Summary updated