Multi-version Machine Learning and Rejuvenation for Resilient Perception in Safety-critical Systems

Machine learning (ML) has become a crucial component in safety-critical systems, such as those used in autonomous vehicle perception. However, the correctness and, therefore, the safety of these systems can be compromised by out-of-distribution data, accidental faults, and security breaches. This paper investigates using a replicated ML architecture to mitigate the risks associated with complex single-points-of-failure. Additionally, it explores the application of rejuvenation to sustain healthy majorities when facing persistent threats. We evaluate the output reliability of the proposed architecture in two case studies: traffic sign detection and perception for autonomous driving. We adopt models and reliability functions, validating our findings using realistic data sets and fault injection experiments. We also evaluate driving safety using the proposed architecture in the CARLA simulator. Our results show that our models can present a good generalization and multi-version ML with proactive rejuvenation can improve correctness and, thus, safety despite faults and cyberattacks.