Architectural robustness in lane detection systems: Geometric constraints and adversarial stability

This study introduces a rule-augmented CNN with spectral normalization that embeds geometric priors directly into the network architecture, evaluated on the TuSimple (highway) and CULane (complex urban) datasets. Robustness is measured as 𝜌 = I o U a t t a c k / I o U c l e a n , where 𝜌 = 1 . 0 denotes no degradation. Across three random seeds, the rule-augmented model improves clean IoU by ＋101% on TuSimple and ＋136% on CULane over a standard CNN, with the larger gain on the more complex urban scenes. Benchmarking against established architectures under a common training protocol, an optimized version of the model attains clean IoU statistically on par with SCNN and exceeds a LaneNet-style baseline, while remaining the only model to provide certified 𝐿 2 robustness (radii 0.307/0.299, agreement > 99%). Critically, although accuracy-based metrics record near-perfect robustness, an IoU-based evaluation reveals that iterative PGD attacks collapse the rule-augmented model ( 𝜌 P G D ≈ 0 . 0 0 1 ) more severely than the standard CNN ( 𝜌 P G D ≈ 0 . 1 4 ): its richer multi-modal input creates a larger adversarial attack surface. This divergence underscores the importance of IoU-based robustness metrics over accuracy-based proxies for imbalanced segmentation tasks and establishes architectural principles for certifiably robust perception in autonomous vehicles.