Defense-in-Depth at the Perception-Reasoning Interface of LLM-Centric Agentic UAV Swarms
Five-layer defense-in-depth protects LLM-controlled UAV swarm scheduling from falsified sensor reports, with closed-form detection boundaries validated in thirty simulations.
The paper implements and evaluates defenses at the perception-reasoning interface of LLM-centric agentic UAV swarms, where an adversary manipulating structured sensor reports can redirect the swarm without touching model weights or the UAV. Five layers check report provenance, physical admissibility, consistency with swarm geometry and service history, schedule starvation, and fall back to a deterministic scheduler ignoring suspect input. Closed-form distortion bounds for the three input-side layers, fixed from deployment parameters before attack data collection, match measured boundaries across thirty matched simulation runs. Replacing rejected reports with the last accepted one prevents adversary control of the schedule but raises cumulative cost by 79% and 74% for the two detectors, while the safety check detects no attacks yet cuts attack-induced cost by 37.5%.
- Adversary can redirect LLM-scheduled UAV swarms by falsifying sensor reports only
- Five layers check provenance, physical admissibility, geometry/history consistency, starvation, deterministic fallback
- Closed-form detection boundaries predicted from deployment parameters match measurements in thirty runs
- Rejecting reports blocks adversary control but increases cumulative cost 74-79% versus undefended system
- Safety check detects no attacks yet reduces attack-induced cost by 37.5%
From the original article
1 min read at arxiv.orgLarge Language Models (LLMs) increasingly support Uncrewed Aerial Vehicle (UAV) swarm operations such as data collection scheduling, where the model reads structured sensor reports and decides which sensors to visit. An adversary who quietly manipulates those reports can redirect the swarm without modifying the model weights or the UAV.
By Mohammadhossein Homaei · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.