Overview
If communicated vehicle states are adversarial or stale, can relational trust improve cooperative control—and what protects the platoon when learned trust is wrong?
Attack detection is not the final cyber-physical outcome. A detector can score well while closed-loop control still violates gaps; a conservative sensor-only controller can remain safe while losing cooperative performance.
Contribution
What was built—and what I did.
SafeGraph-CAV composes edge-level trust estimation, physics-guided features, trust-weighted CACC, and a separate runtime intervention layer, then evaluates the full chain across attacks, held-out seeds, conventional and learned baselines, paired statistics, and SUMO/TraCI.
Developed the project from the initial architecture through the threat model, attack and baseline matrix, trust/control integration, runtime shield, lightweight and SUMO evaluation, statistical validation, diagnostics, reproducibility work, and submission package.
Selected verified evidence
Results with their context attached.
Shielded SafeGraph-CACC versus raw CACC in the principal held-out attacked benchmark.
Shielded SafeGraph-CACC versus chi-square CACC in the same benchmark.
Unsafe samples / collisions / crash penetration for the shielded method in the principal held-out benchmark.
Python benchmark; 1.180 ms p95 and below the nominal 10/20/50 Hz budgets tested.
- Clean RMS remained near raw CACC: approximately 0.252 m, with minimum gap near 13.907 m.
- Shielded SafeGraph-CACC achieved RMS 1.4587 and minimum gap 12.4728 across the principal attacked averages.
- SUMO false-data-injection runs reproduced the prevention pattern, but often through sustained fallback behaviour.
- Trust AUC stayed near 0.99 through moderate delay, then fell to 0.8420 with RMS 13.8890 at the five-step delay stress.
Failures and reversals
Negative results belong in the record.
- An intended communication-noise sensitivity sweep did not propagate its configuration and was excluded from the evidence base.
- Under SUMO FDI-position, unshielded trust control had extreme variance and could be worse than raw CACC in RMS.
- Severe delay and packet absence remained difficult; safe fallback did not restore cooperative tracking quality.
Limitations
The boundary of the claim.
- 01No formal string-stability proof or certified barrier/invariant-set guarantee was derived.
- 02SUMO is not a high-fidelity vehicle-dynamics or real-vehicle validation platform.
- 03The threat model assumes trustworthy onboard sensing and excludes direct actuator compromise.
- 04Large shield activation counts under severe FDI show that fallback can dominate cooperative operation.
Repository
Code and reproducibility surface.
- Modular communication, trust, control, and evaluation packages
- Held-out seed and matched attack protocol
- Resolved baselines, statistical summaries, trajectories, and heatmaps
- Lightweight simulator plus independent SUMO/TraCI execution
Related research log