New SECURE framework addresses critical instability in AI collision prediction models used for autonomous vehicle safety

State-of-the-art accident anticipation models like CRASH show significant prediction instability when exposed to minor input perturbations, raising safety concerns

SECURE framework introduces a multi-objective training methodology that enforces robustness through consistency and stability in both prediction and latent feature spaces

The approach fine-tunes baseline models by minimizing divergence from reference models and penalizing sensitivity to adversarial perturbations

Framework evaluated on DAD and CCD datasets to demonstrate improved reliability for safety-critical autonomous driving applications