Safety is paramount in autonomous vehicles (AVs). Auto manufacturers have
spent millions of dollars and driven billions of miles to prove AVs are safe.
However, this is ill-suited to answer: what happens to an AV if its data are
adversarially compromised? We design a framework built on security-relevant
metrics to benchmark AVs on longitudinal datasets. We establish the
capabilities of a cyber-level attacker with only access to LiDAR datagrams and
from them derive novel attacks on LiDAR. We demonstrate that even though the
attacker has minimal knowledge and only access to raw datagrams, the attacks
compromise perception and tracking in multi-sensor AVs and lead to objectively
unsafe scenarios. To mitigate vulnerabilities and advance secure architectures
in AVs, we present two improvements for security-aware fusion — a
data-asymmetry monitor and a scalable track-to-track fusion of 3D LiDAR and
monocular detections (T2T-3DLM); we demonstrate that the approaches
significantly reduce the attack effectiveness.