Physics-Informed Digital-Twin Evidence Fusion for Resilient Cloud Control of UAV Under Compound Cyberattacks

Cloud control of unmanned aerial vehicles (UAVs) turns communication attacks into physical tracking and safety risks. This paper develops a digital-twin-assisted architecture that fuses observable packet provenance, calibrated innovation, physical-envelope evidence, and temporal anomaly memory. A four-mode supervisor selects nominal or conservative edge/cloud LQR, protected local LQR, or a saturation-aware nonlinear recovery controller. The implementation uses a protected onboard Kalman estimator, removes attack-label leakage, propagates delayed packets causally, separates transmitted and applied commands, and isolates random streams for paired comparisons. A 48-vertex common quadratic Lyapunov certificate covers the three LQR modes within a verified component-wise saturation sector; estimator-aware trajectory headroom yields conditional ellipsoid and magnitude-duration results, while nonlinear recovery is evaluated empirically. The certificate is conditional on the stated linear-mode, saturation-sector, disturbance, and estimator-error assumptions; it is not an end-to-end guarantee for the attacked nonlinear UAV. Nine policies are compared on six reference families under five-episode compound attacks. Across 120 paired nonlinear proposed-policy trials, trajectory-wise mean position RMSE is 0.125-0.129 m, no safety violation or mission termination occurs, and 0.360-0.369 of cloud authority is retained; undefended cloud control terminates every paired trial. The pure evidence alarm recognizes 0.99 of scheduled events, while the protected precursor guard is not credited as cyber detection. Paired comparisons show lower RMSE than LocalOnly and resource-matched Huber and packet-residual supervisors, while the nondeployable oracle remains better. The result is a quantified operating envelope rather than universal resilience.