Estimating the Impact of Soft Errors on AI-Based Perception in Automotive

Artificial intelligence (AI) is boosting the development of complex functionalities in the automotive domain, paving the way for autonomous decision-making capabilities of vehicles. However, integrating such complex functionalities at the system level is challenging due to the short time to market with limited development, verification, and validation periods. Moreover, automotive applications require compliance with strict safety regulations, which demands effective strategies to ensure timely development while allowing thorough dependability evaluations. This work addresses the system-level reliability of automotive applications by proposing a strategy named TIARA (Two-step IntegrAted Reliability Assessment) to assess the impact of hardware faults (modeled as soft errors) on AI-based perception tasks. Our strategy allows for the exploration and evaluation of algorithms, driving controllers, and critical operational scenarios, considering the effects of faults affecting the hardware up to the vehicle level. To tame the huge computational complexity, TIARA first exploits a static analysis to estimate the fault vulnerabilities and identifies the most critical code blocks. Then, the most critical application blocks (susceptible to faults from hardware) are targeted to evaluate the fault's impact on the vehicle dynamics and determine their effect at the system level, combining perception, control, and driving features. Our experimental results, obtained on a test case corresponding to a Lane Centering Assistance application based on the YoloP model and under different driving scenarios, show the feasibility and effectiveness of the proposed approach.