Reading Note
The foundational control background was organized around constant-time-headway platooning. For follower i, the desired spacing was defined as d_i* = d0 + h v_i, and the spacing error compared the actual predecessor gap against that desired spacing. The lightweight vehicle model used position, velocity, and first-order acceleration lag. This model was deliberately simple enough to isolate the effect of corrupted communication and the proposed defense.
Classical work by Swaroop and Hedrick and later work on Lp string stability established why disturbance propagation matters in a platoon. Practical CACC literature showed why upstream velocity and acceleration are useful: communication allows a follower to respond before a change becomes locally visible. This benefit also creates the attack surface. A malicious or stale acceleration message can enter the feedback loop as if it were a physical measurement.
Security-oriented reading covered false-data injection, spoofing, replay, packet loss, delay, and hybrid attacks. A particularly relevant line used onboard sensing and chi-square consistency tests to weight communicated data. This became a baseline rather than the final design because pairwise statistical consistency alone does not exploit relational context across the communication graph.
Technical Essay
The project scope was narrowed to longitudinal platooning. Lane changes, cut-ins, high-fidelity tire models, powertrain detail, actuator compromise, and corrupted onboard ranging were excluded from the core threat model. The adversary would corrupt V2V messages while local physical sensing remained noisy but available.
This restriction was not an attempt to claim a complete autonomous-driving defense. It was necessary to make one mechanism experimentally interpretable: when communication is attacked, can graph-based information selection plus a physical command shield preserve the safety of longitudinal control?
The intended contribution was also refined. Equations for a GNN-PINN trust score alone would be weak. A credible paper required a complete chain: system model, attack model, trust feature definition, training objective, control rule, runtime protection, baselines, simulation, statistical evaluation, and explicit limitations.