A stochastic hybrid automaton for trajectory prediction of CAVs under cyber-induced uncertainty—a signalized intersection study

We develop a Cyber-Aware Hybrid Human–Vehicle Model (CH–HVM) for predicting the longitudinal trajectory of a human-driven connected and automated vehicle (CAV) approaching a signalized intersection when driver-facing infrastructure messages may be spoofed. Cyber-induced uncertainty is defined operationally as a mismatch between the physical signal state and the digital cue observed by the driver, such as a corrupted Red Light Countdown (RLCD) or Signal Phase and Timing (SPaT) message; this mismatch enters the stochastic guard through signal and context features and can alter the modeled probability of stopping versus proceeding. The human–machine interaction between driver, vehicle, and infrastructure information is represented as a stochastic hybrid automaton: a data-driven guard maps contextual features into probabilistic stop/pass decisions, and a structured K -phase braking envelope with a bounded residual term governs anomaly-induced motion. We prove that a sufficient safe-gap condition at a valid stopping-regime entry certifies a minimum stop-line margin at the ego vehicle’s first stop, subject to the one-dimensional dynamics, residual bounds, physical admissibility, and execution-lifetime assumptions. For the separate generic complete-dynamics class, we prove conditional existence and uniqueness with no Zeno behavior. We also derive episode-wise model-risk bounds and a finite-sample population bound for a fixed model evaluated on independent inputs. A high-fidelity driving-simulator experiment with spoofed RLCD messages provides descriptive evidence on braking and low-speed responses in 32 trajectory extracts from this single-vehicle intersection setting. Unlike purely physics-based models, CH–HVM embeds a data-driven representation of cyber-affected human decisions and braking behavior; unlike unconstrained machine learning models, it yields closed-form conditional first-stop certificates and statistically justified fixed-model risk bounds for this single-vehicle intersection scenario under adversarial infrastructure messages.