Improving Robustness for ADS: Transforming Input Distributions to Account for Inference and System State 51 views

Autonomous Driving Systems (ADSs) increasingly rely on deep neural networks (DNNs). As these systems grow in autonomy, responsibility, and deployment scale, the consequences of their failures become more severe. In this dissertation, we show that such failures often stem from unanticipated shifts in the input distribution -- caused either by long-tail driving scenarios or by the evolving state of the system itself -- that can degrade the robustness of DNN-based components. To address these vulnerabilities, we present a suite of techniques that systematically identify, characterize, and mitigate failures caused by input distribution shifts. First, we introduce PreFixer, a method that adapts DNNs to hardware changes -- such as sensor swaps -- by automatically generating transformations that compensate for discrepancies between training and deployment data. Second, we extend adversarial testing to better reflect real-world conditions through DeepManeuver, which models the compounding effects of perturbations over time and integrates system-level feedback into the attack loop. Finally, we develop NaturalADV, a framework for generating adversarial perturbations that are both effective and perceptually natural. The study on NaturalADV shows that varying definitions of naturalness and optimization techniques can have significant impact on perturbation strength and can preserve naturalness from a benign natural patch and still produce strong perturbations that enact system-level failures. Together, these contributions advance the state of the art in testing and enhancing ADS robustness for learning-enabled systems and provide a foundation for more robust autonomy in the safety-critical ADS domain.

PHD (Doctor of Philosophy)

Robotics; Software Engineering; Neural Networks; System Testing; Adversarial Attacks; Autonomous Vehicles; Autonomy

English

2025-07-28