Progress Update
By this point the research direction was explicitly described as physics-guided graph trust, runtime shielding, attack detection/prevention, and platoon safety. The project was no longer framed as a standalone intrusion detector.
Technical Essay
The communication topology was represented as a directed graph G_t = (V, E_t). A directed edge (j, i) represented a packet sent by vehicle j and used by vehicle i. This representation was important because trust is not simply a permanent property of a vehicle. It can vary by sender-receiver relation, time, packet freshness, attack, and surrounding graph context.
The planned edge feature vector combined:
- the communicated sender state;
- the receiver or ego state;
- relative position, velocity, and acceleration features;
- a short-horizon kinematic or physics residual;
- chi-square/statistical consistency evidence;
- freshness, availability, and channel metadata.
The model output w_hat_ij,t in [0,1] would be interpreted as a learned reliability estimate. Trust values would be normalized across candidate predecessor sources and used inside the cooperative control term. A small trust floor would prevent numerical collapse when all messages were assigned very low scores.
The runtime shield would monitor physical inter-vehicle gap and time to collision. If either crossed a configured threshold, the trust-weighted command would be replaced by a conservative fallback command. This allowed the paper to make a scoped empirical claim: learned trust improves the selection of cooperative information; the runtime shield prevents the final controller from relying on trust alone when the physical state becomes dangerous.
Experiment Note
An early prototype run produced approximately 0.3538 m RMS spacing error, 13.8369 m minimum gap, zero collisions, zero unsafe-gap rate, and 0.2431 m/s mean velocity error over 300 steps. The same run did not satisfy the preliminary empirical string-stability diagnostic. These values were treated only as an initial systems check. They were not sufficient for a paper because they came from a representative configuration rather than a held-out multi-seed benchmark.