Defense-in-Depth at the Perception-Reasoning Interface of LLM-Centric Agentic UAV Swarms
What it is
This paper proposes and evaluates a five-layer defense-in-depth system for LLM-driven UAV swarms to protect against adversarial manipulation of sensor reports. It implements checks for provenance, physical admissibility, geometric consistency, and starvation, falling back to a deterministic scheduler if inputs are suspect.
The problem it solves
Adversaries can manipulate structured sensor reports to redirect LLM-controlled UAV swarms without modifying model weights or hardware, a vulnerability that existing architectural defenses have rarely implemented or evaluated.
What you could use it for
- Securing LLM-based decision-making in autonomous drone swarms
- Preventing adversarial redirection of UAV data collection schedules
- Implementing robust input validation for agentic AI systems
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.