Experiment Note
The held-out benchmark established the main result. In clean operation, raw CACC, unshielded SafeGraph-CACC, and shielded SafeGraph-CACC behaved almost identically, with RMS spacing error near 0.252 m and minimum gap near 13.907 m. This showed that the learned trust and shield did not materially degrade normal cooperative behavior.
Across attacked scenarios, the principal averages were:
- Raw CACC: RMS 5.7730, minimum gap 3.8652, unsafe count 87.86, collisions 58.98, crash metric 862.44.
- Chi-square CACC: RMS 5.1312, minimum gap -0.7691, unsafe count 95.69, collisions 76.00, crash metric 1481.23.
- Residual Trust CACC: RMS 1.5257, minimum gap 9.2829, unsafe count 32.77, collisions 22.60, crash metric 212.17.
- Unshielded SafeGraph-CACC: RMS 1.5261, minimum gap 11.2978, unsafe count 13.93, collisions 7.46, crash metric 15.71.
- Shielded SafeGraph-CACC: RMS 1.4587, minimum gap 12.4728, zero unsafe samples, zero collisions, and zero crash metric.
- Oracle Trust CACC: RMS 0.2851, minimum gap 13.9048, zero unsafe samples, zero collisions, and zero crash metric.
Relative to raw CACC, Shielded SafeGraph-CACC reduced attacked-scenario RMS spacing error by 74.73 percent. Relative to chi-square CACC, the reduction was 71.57 percent. Relative to unshielded SafeGraph, the shielded version improved RMS by 4.41 percent while removing observed collisions and crash penetration.
Technical Essay
The most important finding was not that graph trust alone solved the problem. It did not. Unshielded SafeGraph substantially improved trust quality and tracking but still allowed residual unsafe events. The paper's strongest mechanism-level conclusion became: graph trust improves cooperative information selection; runtime shielding converts that improvement into empirically safe closed-loop behavior in the evaluated attack model.
The oracle gap also remained large. This showed that there was substantial room for better trust calibration, delay handling, and missing-information reconstruction.