Autonomous Vehicle Resilience
The emergence of ML-driven systems and their ubiquitous adoption means that their trustworthiness and dependability are now of paramount importance. Autonomous vehicle (AV) technologies are the perfect examples of such systems and are advertised to be transformative, with the potential to improve traffic congestion, safety, productivity, and comfort. With the increasing popularity and ubiquitous deployment of AVs on public roads, dependability, and trustworthiness have increasingly become critical requirements for public acceptance and adoption. A trustworthy system must be functionally correct, robust, safe, resilient, privacy-preserving, and secure.
We have developed a suite of ML-driven techniques and associated tools for assessing resilience and security of AVs:
- Field-Data Assessment: We have built LogDriver to parse and statistically analyze disengagement and accident reports obtained from public DMV databases, thereby providing a method to investigate the causes, dynamics, and impacts of AV failures in the wild.
- Fault Injection: We have built fault-selection and -injection techniques, and tools for end-to-end
reliability assessment of AVs.
- We have built AVFI, a fault injector that targets sensor fault and failure models.
- We have built BDLFI, a fault injector for neural network training and inference, which uses Bayesian deep learning to model fault propagation.
- We have built BFI, an intelligent resiliency assessment tool that can automatically identify situations and faults that will likely lead to safety violations. The BFI relies on an ML-based fault selection engine for causal and counterfactual reasoning about the system state in terms of safety under a fault scenario.
- Adversarial Attacks: We created RoboTack, an intelligent malware that can disguise attacks as accidental/random to evade detection, yet cause serious safety/reliability incidents (e.g., an accident of an autonomous vehicle). The attack targets the use of adversarial examples in the multiple object tracking process of the AV’s perception pipeline.