A Digital Twin-Based Cybersecurity Architecture for FDI Anomaly Detection and Isolation in Next Generation Robotic Vehicles

Abstract

This work proposes a four-layer resilient architecture for trajectory tracking of a mobile vehicle subject to false data injection (FDI) attacks on its pose sensor. The architecture combines a channel-wise statistical detector, a kinematic digital twin that predicts the vehicle’s state independently of the incoming sensor readings, and a supervisory controller that shifts smoothly between trusting the sensor and relying on the digital twin, depending on whether an attack is detected. This design allows the vehicle to keep tracking its intended path during an attack and to return to normal, sensor based operation afterward, without needing to be retuned for each new trajectory or attack type. The architecture is evaluated over four trajectories and nine attack types, covering 180 simulation runs with realistic measurement noise, and reaches a pass rate of 100%, with no false activations in attack free conditions. We further compare the proposed architecture against three baseline detectors evaluated under the same conditions, including a Kalman filter paired with a simple accept or reject rule, a recent set based detector, and a classical statistical residual test. All protected methods achieve a 100% pass rate with comparable tracking accuracy (mean RMSE 0.052–0.059 m); notably, the proposed architecture yields the highest mean RMSE (0.0585 m) among the four protected methods, as its conservative CUSUM thresholds and consistency-gated recovery prioritize false-alarm avoidance over fastest re-engagement. The proposed architecture produces smoother control commands than the baselines, a property that matters for real actuators. A control experiment confirms that this smoothness advantage is attributable specifically to the sigmoid-based blending in the handover layer, rather than to the detection or estimation layers of the architecture. The results show that combining statistical detection with a digital twin based supervisory controller can maintain both accuracy and control smoothness under sensor attacks, without requiring per scenario retuning.