Progress Update
The main benchmark was organized around disjoint training and evaluation seeds. SafeGraph was trained on seeds 0-19 and evaluated on held-out seeds 100-119. Attack-method pairs were matched by seed so that comparisons could later use paired statistical tests.
Experiment Note
The metric suite was broadened to prevent one attractive aggregate number from hiding unsafe behavior. It included:
- RMS and maximum spacing error;
- velocity error;
- minimum physical gap;
- unsafe-gap sample count;
- TTC violation count;
- collision count;
- crash-penetration metric;
- trust ROC-AUC and F1;
- false-trust rate;
- false-alarm rate;
- shield activation count;
- runtime latency.
Technical Essay
The metric design reflects an important cyber-physical principle: attack classification accuracy is not the final outcome. A detector can obtain strong ROC-AUC while the closed-loop controller still violates safety constraints. Conversely, a conservative sensor-only controller can be safe while sacrificing cooperative tracking performance. The evaluation therefore followed the entire causal chain from trust quality to control accuracy to physical safety.
The oracle baseline was included as an upper bound, not a deployable method. It answered whether poor performance was caused by the basic control formulation or by imperfect trust estimation.