Preprint tests defenses against sensor-report attacks on AI drone swarms

Researchers report in a preprint that defenses against manipulated sensor reports prevented an attacker from controlling an AI-directed drone swarm's schedule, but increased cumulative cost. Replacing rejected input with the most recent accepted report increased cost by 79% and 74% for two detectors, respectively, compared with the undefended system. The framework checks report provenance, physical plausibility, consistency with swarm geometry and service history, and whether schedules neglect sensors. It also provides a deterministic fallback scheduler that ignores suspect input. Across thirty matched simulation runs, predicted and measured detection boundaries agreed. A separate safety check detected no attacks but reduced attack-induced cost by 37.5%.