SHIELD: Secure Human-Machine Interaction with Evidential Learning and Dynamic Trust for Drone Swarm Control

Drone swarms face a dual challenge: human operators cannot effectively supervise more than a dozen Unmanned Aerial Vehicles (UAVs) before situational awareness degrades, yet fully autonomous coordination remains vulnerable to cyber-physical attacks such as GPS spoofing, communication jamming, Byzantine manipulation, replay, malware injection, and network intrusion. We present SHIELD, a framework for secure human--swarm teaming built on a single unifying insight: evidential vacuity, derived from Dirichlet-based perception, jointly drives inter-agent trust modulation, intrusion detection, and human escalation within a decentralized control loop. A dynamic trust graph filters unreliable agents through behavioral consistency and quarantine, while an embedded intrusion response system autonomously handles detection, containment, and recovery. Operator involvement is reserved for genuinely ambiguous cases via an event-triggered binary interface that minimizes workload. In the considered waypoint-surveillance simulation scenarios with six attack classes and swarms of 8--32 UAVs, SHIELD achieves up to 60% relative improvement in mission success, detects over 80% of compromised agents with under 15% false positives, and reduces human interventions by 53% compared to baselines. Ablation and scalability studies confirm complementary module contributions and graceful performance degradation.