Tonmoay Deb

Commercial drones weaponized by non-state actors achieve attack success rates above 70%, yet existing counter-drone systems focus on detection rather than threat assessment and legally compliant response. This dissertation develops four interconnected systems for autonomous urban drone defense:

DEWS (Drone Early Warning System) classifies drone trajectories as threatening or benign based on initial flight segments. Evaluated on 349 real-world trajectories provided by the Dutch Police over eight months in The Hague, DEWS achieves F1-scores exceeding 0.80 within 30 seconds of initial observation.

STATE (Safe and Threatening Adversarial Trajectory Engine) addresses data scarcity through conditional generative adversarial networks that synthesize threat-conditioned drone trajectories based on geographic context.

POSS (Pareto-Optimal Status Sets) formalizes multi-objective decision-making under legal constraints using deontic logic and Pareto optimization, ensuring autonomous agents never select legally impermissible actions.

GUARDIAN (Governance-Unified Aerial Reinforcement-Defense In Accordance with Norms) coordinates defensive drone swarms using multi-agent reinforcement learning while maintaining strict legal compliance through POSS-based action masking. A key finding is that incorporating legal constraints can actually improve defensive performance when defenders are outnumbered, challenging assumptions that compliance necessarily compromises tactical effectiveness.