RVPLAYER: Robotic Vehicle Forensics by Replay with What-if Reasoning

Hongjun Choi (Purdue University), Zhiyuan Cheng (Purdue University), Xiangyu Zhang (Purdue University)

Robotic vehicle (RV) attack forensics identifies root cause of an accident.
Reproduction of accident and reasoning about its causation are critical steps in
the process. Ideally, such investigation would be performed in real-world field tests by faithfully regenerating the environmental conditions and varying the different factors to understand causality. However, such analysis is extremely expensive and in many cases infeasible due to the difficulties of enforcing physical conditions. Existing RV forensics techniques focus on faithful accident reproduction in simulation and hence lack the support of causality
reasoning. They also entail substantial overhead.
We propose RVPLAYER, a system for RV forensics. It supports replay with what-if reasoning inside simulator (e.g., checking if an accident can be avoided by changing some control parameter, code, or vehicle states). It is a low-cost replacement of the expensive field test based forensics. It features an efficient demand-driven adaptive logging method capturing non-deterministic physical conditions, and a novel replay technique supporting various replay policies that selectively enable/disable information
during replay for root cause analysis. Our evaluation on 6 RVs (4 real and 2 virtual), 5 real-world auto-driving traces, and 1194 attack instances of various kinds reported in the literature shows that it can precisely pinpoint the root causes of these attacks without false positives. It has only 6.57% of the overhead of a simple logging design.