A particularly challenging problem in AI safety is providing guarantees on the behavior of high-dimensional autonomous systems. Verification approaches centered around reachability analysis fail to scale, and purely statistical approaches are constrained by the distributional assumptions about the sampling process. Instead, we pose a distributionally robust version of the statistical verification problem for black-box systems, where our performance guarantees hold over a large family of distributions. This paper proposes a novel approach based on uncertainty quantification using concepts from imprecise probabilities. A central piece of our approach is an ensemble technique called Imprecise Neural Networks, which provides the uncertainty quantification. Additionally, we solve the allied problem of exploring the input set using active learning. The active learning uses an exhaustive neural-network verification tool Sherlock to collect samples. An evaluation on multiple physical simulators in the openAI gym Mujoco environments with reinforcement-learned controllers demonstrates that our approach can provide useful and scalable guarantees for high-dimensional systems.

Imitation Learning (IL) has shown significant promise in autonomous driving, but its performance heavily depends on the quality of training data. Noisy or corrupted sensor inputs can degrade learned policies, leading to unsafe behavior. This paper presents an unsupervised anomaly detection approach to automatically filter out abnormal images from driving datasets, thereby enhancing IL performance. Our method leverages a Convolutional Autoencoder with a novel latent reference loss, which forces abnormal images to reconstruct with higher errors than normal images. This enables effective anomaly detection without requiring manually labeled data. We validate our approach on the realistic DonkeyCar autonomous racing platform, demonstrating that filtering videos significantly improves IL policies, as measured by a 25-40% reduction in cross-track error. Compared to baseline and ablation models, our method achieves superior anomaly detection across three real-world video corruptions: collision-based occlusions, transparent obstructions, and raindrop interference. The results highlight the effectiveness of unsupervised video anomaly detection in improving the robustness and performance of IL-based autonomous control.Video: https://youtu.be/RjJ3nZR6RQ

Vision-based control relies on accurate perception to achieve robustness. However, image distribution changes caused by sensor noise, adverse weather, and dynamic lighting can degrade perception, leading to suboptimal control decisions. Existing approaches, including domain adaptation and adversarial training, improve robustness but struggle to generalize to unseen corruptions while introducing computational overhead. To address this challenge, we propose a real-time image repair module that restores corrupted images before they are used by the controller. Our method leverages generative adversarial models, specifically CycleGAN and pix2pix, for image repair. CycleGAN enables unpaired image-to-image translation to adapt to novel corruptions, while pix2pix exploits paired image data when available to improve the quality. To ensure alignment with control performance, we introduce a control-focused loss function that prioritizes perceptual consistency in repaired images. We evaluated our method in a simulated autonomous racing environment with various visual corruptions. The results show that our approach significantly improves performance compared to baselines, mitigating distribution shift and enhancing controller reliability.

Medical cyber-physical systems (MCPS) are increasingly adopting learning-enabled components (LECs) to enhance their decision-making capabilities. Due to the safety-critical nature of MCPS, these systems must maintain high performance on both expected and unexpected input data. Therefore, ensuring the robustness of LE-MCPS is crucial for their successful deployment. Existing research predominantly focuses on robustness to synthetic adversarial examples , crafted by adding imperceptible perturbations to clean input data. However, these synthetic adversarial examples do not accurately reflect the most challenging real-world scenarios, especially in the context of healthcare data. Consequently, robustness to synthetic adversarial examples may not necessarily translate to robustness against naturally occurring adversarial examples . We propose a method to evaluate the robustness of LE-MCPS to natural adversarial examples. The method curates naturally adversarial datasets leveraging probabilistic labels obtained from automated weakly supervised labeling which combines noisy and cheap-to-obtain labeling heuristics. Based on these labels, the method adversarially orders the input data and uses this ordering to construct a sequence of increasingly adversarial datasets for assessing robustness. Our evaluation on six MCPS case studies and two non-medical case studies demonstrates (1) the efficacy and statistical validity of our approach to generating naturally adversarial datasets and (2) the utility of our robustness evaluation in classifying robust and non-robust LE-MCPS.

Abstract Autonomous systems are increasingly implemented using end-to-end learning-based controllers. Such controllers make decisions that are executed on the real system, with images as one of the primary sensing modalities. Deep neural networks form a fundamental building block of such controllers. Unfortunately, the existing neural-network verification tools do not scale to inputs with thousands of dimensions—especially when the individual inputs (such as pixels) are devoid of clear physical meaning. This paper takes a step towards connecting exhaustive closed-loop verification with high-dimensional controllers. Our key insight is that the behavior of a high-dimensional vision-based controller can be approximated with several low-dimensional controllers. To balance the approximation accuracy and verifiability of our low-dimensional controllers, we leverage the latest verification-aware knowledge distillation. Then, we inflate low-dimensional reachability results with statistical approximation errors, yielding a high-confidence reachability guarantee for the high-dimensional controller. We investigate two inflation techniques—based on trajectories and control actions—both of which show convincing performance in three OpenAI gym benchmarks.

Classification of clinical alarms is at the heart of prioritization, suppression, integration, postponement, and other methods of mitigating alarm fatigue. Since these methods directly affect clinical care, alarm classifiers, such as intelligent suppression systems, need to be evaluated in terms of their sensitivity and specificity, which is typically calculated on a labeled dataset of alarms. Unfortunately, the collection and particularly labeling of such datasets requires substantial effort and time, thus deterring hospitals from investigating mitigations of alarm fatigue. This article develops a lightweight method for evaluating alarm classifiers without perfect alarm labels. The method relies on probabilistic labels obtained from data programming—a labeling paradigm based on combining noisy and cheap-to-obtain labeling heuristics. Based on these labels, the method produces confidence bounds for the sensitivity/specificity values from a hypothetical evaluation with manual labeling. Our experiments on five alarm datasets collected at Children’s Hospital of Philadelphia show that the proposed method provides accurate bounds on the classifier’s sensitivity/specificity, appropriately reflecting the uncertainty from noisy labeling and limited sample sizes.

Runtime monitoring is a vital part of safety-critical systems. However, early stage assurance of monitoring quality is currently limited: it relies either on complex models that might be inaccurate in unknown ways or on data that would only be available once the system has been built. To address this issue, we propose a compositional framework for modeling and analysis of noisy monitoring systems. Our novel 3-value detector model uses probability spaces to represent atomic (noncomposite) detectors, and it composes them into a temporal logic-based monitor. The error rates of these monitors are estimated by our analysis engine, which combines symbolic probability algebra, independence inference, and estimation from labeled detection data. Our evaluation on an autonomous underwater vehicle found that our framework produces accurate estimates of error rates while using only detector traces, without any monitor traces. Furthermore, when data are scarce, our approach shows higher accuracy than noncompositional data-driven estimates from monitor traces. Thus, this article enables accurate evaluation of logical monitors in early design stages before deploying them.

Developing cyber-physical systems involves multiple engineering domains, e.g., timing, logical correctness, thermal resilience, and mechanical stress. In today's industrial practice, these domains rely on multiple analyses to obtain and verify critical system properties. Domain differences make the analyses abstract away interactions among themselves, potentially invalidating the results. Specifically, one challenge is to ensure that an analysis is never applied to a model that violates the assumptions of the analysis. Since such violation can originate from the updating of the model by another analysis, analyses must be executed in the correct order. Another challenge is to apply diverse analyses soundly and scalably over models of realistic complexity. To address these challenges, we develop an analysis integration approach that uses contracts to specify dependencies between analyses, determine their correct orders of application, and specify and verify applicability conditions in multiple domains. We implement our approach and demonstrate its effectiveness, scalability, and extensibility through a verification case study for thread and battery cell scheduling.