Insup Lee

Kaustubh Sridhar, Souradeep Dutta, James Weimer et al.

Deep neural networks have emerged as the workhorse for a large section of robotics and control applications, especially as models for dynamical systems. Such data-driven models are in turn used for designing and verifying autonomous systems. They are particularly useful in modeling medical systems where data can be leveraged to individualize treatment. In safety-critical applications, it is important that the data-driven model is conformant to established knowledge from the natural sciences. Such knowledge is often available or can often be distilled into a (possibly black-box) model. For instance, an F1 racing car should conform to Newton's laws (which are encoded within a unicycle model). In this light, we consider the following problem - given a model $M$ and a state transition dataset, we wish to best approximate the system model while being a bounded distance away from $M$. We propose a method to guarantee this conformance. Our first step is to distill the dataset into a few representative samples called memories, using the idea of a growing neural gas. Next, using these memories we partition the state space into disjoint subsets and compute bounds that should be respected by the neural network in each subset. This serves as a symbolic wrapper for guaranteed conformance. We argue theoretically that this only leads to a bounded increase in approximation error; which can be controlled by increasing the number of memories. We experimentally show that on three case studies (Car Model, Drones, and Artificial Pancreas), our constrained neurosymbolic models conform to specified models (each encoding various constraints) with order-of-magnitude improvements compared to the augmented Lagrangian and vanilla training methods. Our code can be found at: https://github.com/kaustubhsridhar/Constrained_Models

Radoslav Ivanov, James Weimer, Rajeev Alur et al.

This paper presents Verisig, a hybrid system approach to verifying safety properties of closed-loop systems using neural networks as controllers. Although techniques exist for verifying input/output properties of the neural network itself, these methods cannot be used to verify properties of the closed-loop system (since they work with piecewise-linear constraints that do not capture non-linear plant dynamics). To overcome this challenge, we focus on sigmoid-based networks and exploit the fact that the sigmoid is the solution to a quadratic differential equation, which allows us to transform the neural network into an equivalent hybrid system. By composing the network's hybrid system with the plant's, we transform the problem into a hybrid system verification problem which can be solved using state-of-the-art reachability tools. We show that reachability is decidable for networks with one hidden layer and decidable for general networks if Schanuel's conjecture is true. We evaluate the applicability and scalability of Verisig in two case studies, one from reinforcement learning and one in which the neural network is used to approximate a model predictive controller.

Sangdon Park, Edgar Dobriban, Insup Lee et al. · gatech

Uncertainty quantification is a key component of machine learning models targeted at safety-critical systems such as in healthcare or autonomous vehicles. We study this problem in the context of meta learning, where the goal is to quickly adapt a predictor to new tasks. In particular, we propose a novel algorithm to construct \emph{PAC prediction sets}, which capture uncertainty via sets of labels, that can be adapted to new tasks with only a few training examples. These prediction sets satisfy an extension of the typical PAC guarantee to the meta learning setting; in particular, the PAC guarantee holds with high probability over future tasks. We demonstrate the efficacy of our approach on four datasets across three application domains: mini-ImageNet and CIFAR10-C in the visual domain, FewRel in the language domain, and the CDC Heart Dataset in the medical domain. In particular, our prediction sets satisfy the PAC guarantee while having smaller size compared to other baselines that also satisfy this guarantee.

Ramneet Kaur, Kaustubh Sridhar, Sangdon Park et al.

Machine learning models are prone to making incorrect predictions on inputs that are far from the training distribution. This hinders their deployment in safety-critical applications such as autonomous vehicles and healthcare. The detection of a shift from the training distribution of individual datapoints has gained attention. A number of techniques have been proposed for such out-of-distribution (OOD) detection. But in many applications, the inputs to a machine learning model form a temporal sequence. Existing techniques for OOD detection in time-series data either do not exploit temporal relationships in the sequence or do not provide any guarantees on detection. We propose using deviation from the in-distribution temporal equivariance as the non-conformity measure in conformal anomaly detection framework for OOD detection in time-series data.Computing independent predictions from multiple conformal detectors based on the proposed measure and combining these predictions by Fisher's method leads to the proposed detector CODiT with guarantees on false detection in time-series data. We illustrate the efficacy of CODiT by achieving state-of-the-art results on computer vision datasets in autonomous driving. We also show that CODiT can be used for OOD detection in non-vision datasets by performing experiments on the physiological GAIT sensory dataset. Code, data, and trained models are available at https://github.com/kaustubhsridhar/time-series-OOD.

Shuo Li, Sangdon Park, Xiayan Ji et al. · gatech

Accurately detecting and tracking multi-objects is important for safety-critical applications such as autonomous navigation. However, it remains challenging to provide guarantees on the performance of state-of-the-art techniques based on deep learning. We consider a strategy known as conformal prediction, which predicts sets of labels instead of a single label; in the classification and regression settings, these algorithms can guarantee that the true label lies within the prediction set with high probability. Building on these ideas, we propose multi-object detection and tracking algorithms that come with probably approximately correct (PAC) guarantees. They do so by constructing both a prediction set around each object detection as well as around the set of edge transitions; given an object, the detection prediction set contains its true bounding box with high probability, and the edge prediction set contains its true transition across frames with high probability. We empirically demonstrate that our method can detect and track objects with PAC guarantees on the COCO and MOT-17 datasets.

Matthew Cleaveland, Insup Lee, George J. Pappas et al.

Conformal prediction is a statistical tool for producing prediction regions of machine learning models that are valid with high probability. However, applying conformal prediction to time series data leads to conservative prediction regions. In fact, to obtain prediction regions over $T$ time steps with confidence $1-δ$, {previous works require that each individual prediction region is valid} with confidence $1-δ/T$. We propose an optimization-based method for reducing this conservatism to enable long horizon planning and verification when using learning-enabled time series predictors. Instead of considering prediction errors individually at each time step, we consider a parameterized prediction error over multiple time steps. By optimizing the parameters over an additional dataset, we find prediction regions that are not conservative. We show that this problem can be cast as a mixed integer linear complementarity program (MILCP), which we then relax into a linear complementarity program (LCP). Additionally, we prove that the relaxed LP has the same optimal cost as the original MILCP. Finally, we demonstrate the efficacy of our method on case studies using pedestrian trajectory predictors and F16 fighter jet altitude predictors.

Michele Caprio, Souradeep Dutta, Kuk Jin Jang et al.

Uncertainty quantification and robustness to distribution shifts are important goals in machine learning and artificial intelligence. Although Bayesian Neural Networks (BNNs) allow for uncertainty in the predictions to be assessed, different sources of predictive uncertainty cannot be distinguished properly. We present Credal Bayesian Deep Learning (CBDL). Heuristically, CBDL allows to train an (uncountably) infinite ensemble of BNNs, using only finitely many elements. This is possible thanks to prior and likelihood finitely generated credal sets (FGCSs), a concept from the imprecise probability literature. Intuitively, convex combinations of a finite collection of prior-likelihood pairs are able to represent infinitely many such pairs. After training, CBDL outputs a set of posteriors on the parameters of the neural network. At inference time, such posterior set is used to derive a set of predictive distributions that is in turn utilized to distinguish between (predictive) aleatoric and epistemic uncertainties, and to quantify them. The predictive set also produces either (i) a collection of outputs enjoying desirable probabilistic guarantees, or (ii) the single output that is deemed the best, that is, the one having the highest predictive lower probability -- another imprecise-probabilistic concept. CBDL is more robust than single BNNs to prior and likelihood misspecification, and to distribution shift. We show that CBDL is better at quantifying and disentangling different types of (predictive) uncertainties than single BNNs and ensemble of BNNs. In addition, we apply CBDL to two case studies to demonstrate its downstream tasks capabilities: one, for motion prediction in autonomous driving scenarios, and two, to model blood glucose and insulin dynamics for artificial pancreas control. We show that CBDL performs better when compared to an ensemble of BNNs baseline.

Yahan Yang, Soham Dan, Dan Roth et al.

Recently it has been shown that state-of-the-art NLP models are vulnerable to adversarial attacks, where the predictions of a model can be drastically altered by slight modifications to the input (such as synonym substitutions). While several defense techniques have been proposed, and adapted, to the discrete nature of text adversarial attacks, the benefits of general-purpose regularization methods such as label smoothing for language models, have not been studied. In this paper, we study the adversarial robustness provided by various label smoothing strategies in foundational models for diverse NLP tasks in both in-domain and out-of-domain settings. Our experiments show that label smoothing significantly improves adversarial robustness in pre-trained models like BERT, against various popular attacks. We also analyze the relationship between prediction confidence and robustness, showing that label smoothing reduces over-confident errors on adversarial examples.

Shuo Li, Xiayan Ji, Edgar Dobriban et al.

Anomaly detection is essential for preventing hazardous outcomes for safety-critical applications like autonomous driving. Given their safety-criticality, these applications benefit from provable bounds on various errors in anomaly detection. To achieve this goal in the semi-supervised setting, we propose to provide Probably Approximately Correct (PAC) guarantees on the false negative and false positive detection rates for anomaly detection algorithms. Our method (PAC-Wrap) can wrap around virtually any existing semi-supervised and unsupervised anomaly detection method, endowing it with rigorous guarantees. Our experiments with various anomaly detectors and datasets indicate that PAC-Wrap is broadly effective.

Michele Caprio, Yusuf Sale, Eyke Hüllermeier et al.

In their seminal 1990 paper, Wasserman and Kadane establish an upper bound for the Bayes' posterior probability of a measurable set $A$, when the prior lies in a class of probability measures $\mathcal{P}$ and the likelihood is precise. They also give a sufficient condition for such upper bound to hold with equality. In this paper, we introduce a generalization of their result by additionally addressing uncertainty related to the likelihood. We give an upper bound for the posterior probability when both the prior and the likelihood belong to a set of probabilities. Furthermore, we give a sufficient condition for this upper bound to become an equality. This result is interesting on its own, and has the potential of being applied to various fields of engineering (e.g. model predictive control), machine learning, and artificial intelligence.

Souradeep Dutta, Michele Caprio, Vivian Lin et al.

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.

Ramneet Kaur, Xiayan Ji, Souradeep Dutta et al.

As machine learning models continue to achieve impressive performance across different tasks, the importance of effective anomaly detection for such models has increased as well. It is common knowledge that even well-trained models lose their ability to function effectively on out-of-distribution inputs. Thus, out-of-distribution (OOD) detection has received some attention recently. In the vast majority of cases, it uses the distribution estimated by the training dataset for OOD detection. We demonstrate that the current detectors inherit the biases in the training dataset, unfortunately. This is a serious impediment, and can potentially restrict the utility of the trained model. This can render the current OOD detectors impermeable to inputs lying outside the training distribution but with the same semantic information (e.g. training class labels). To remedy this situation, we begin by defining what should ideally be treated as an OOD, by connecting inputs with their semantic information content. We perform OOD detection on semantic information extracted from the training data of MNIST and COCO datasets and show that it not only reduces false alarms but also significantly improves the detection of OOD inputs with spurious features from the training data.

Vivian Lin, Kuk Jin Jang, Souradeep Dutta et al.

Deep neural networks have repeatedly been shown to be non-robust to the uncertainties of the real world, even to naturally occurring ones. A vast majority of current approaches have focused on data-augmentation methods to expand the range of perturbations that the classifier is exposed to while training. A relatively unexplored avenue that is equally promising involves sanitizing an image as a preprocessing step, depending on the nature of perturbation. In this paper, we propose to use control for learned models to recover from distribution shifts online. Specifically, our method applies a sequence of semantic-preserving transformations to bring the shifted data closer in distribution to the training set, as measured by the Wasserstein distance. Our approach is to 1) formulate the problem of distribution shift recovery as a Markov decision process, which we solve using reinforcement learning, 2) identify a minimum condition on the data for our method to be applied, which we check online using a binary classifier, and 3) employ dimensionality reduction through orthonormal projection to aid in our estimates of the Wasserstein distance. We provide theoretical evidence that orthonormal projection preserves characteristics of the data at the distributional level. We apply our distribution shift recovery approach to the ImageNet-C benchmark for distribution shifts, demonstrating an improvement in average accuracy of up to 14.21% across a variety of state-of-the-art ImageNet classifiers. We further show that our method generalizes to composites of shifts from the ImageNet-C benchmark, achieving improvements in average accuracy of up to 9.81%. Finally, we test our method on CIFAR-100-C and report improvements of up to 8.25%.

Souradeep Dutta, Yahan Yang, Elena Bernardis et al.

The performance of machine learning models can significantly degrade under distribution shifts of the data. We propose a new method for classification which can improve robustness to distribution shifts, by combining expert knowledge about the ``high-level" structure of the data with standard classifiers. Specifically, we introduce two-stage classifiers called memory classifiers. First, these identify prototypical data points -- memories -- to cluster the training data. This step is based on features designed with expert guidance; for instance, for image data they can be extracted using digital image processing algorithms. Then, within each cluster, we learn local classifiers based on finer discriminating features, via standard models like deep neural networks. We establish generalization bounds for memory classifiers. We illustrate in experiments that they can improve generalization and robustness to distribution shifts on image datasets. We show improvements which push beyond standard data augmentation techniques.

Luan Viet Nguyen, Gautam Mohan, James Weimer et al.

Model-based design offers a promising approach for assisting developers to build reliable and secure cyber-physical systems (CPSs) in a systematic manner. In this methodology, a designer first constructs a model, with mathematically precise semantics, of the system under design, and performs extensive analysis with respect to correctness requirements before generating the implementation from the model. However, as new vulnerabilities are discovered, requirements evolve aimed at ensuring resiliency. There is currently a shortage of an inexpensive, automated mechanism that can effectively repair the initial design, and a model-based system developer regularly needs to redesign and reimplement the system from scratch. In this paper, we propose a new methodology along with a Matlab toolkit called REAFFIRM to facilitate the model-based repair for improving the resiliency of CPSs. REAFFIRM takes the inputs including 1) an original hybrid system modeled as a Simulink/Stateflow diagram, 2) a given resiliency pattern specified as a model transformation script, and 3) a safety requirement expressed as a Signal Temporal Logic formula, and then outputs a repaired model which satisfies the requirement. The overall structure of REAFFIRM contains two main modules, a model transformation, and a model synthesizer built on top of the falsification tool Breach. We introduce a new model transformation language for hybrid systems, which we call HATL to allow a designer to specify resiliency patterns. To evaluate the proposed approach, we use REAFFIRM to automatically synthesize repaired models for an adaptive cruise control (ACC) system under a GPS sensor spoofing attack, for a single-machine infinite-bus (SMIB) system under a sliding-mode switching attack, and for a missile guidance system under gyroscopes sensor attack.

Pengyuan Lu, Ivan Ruchkin, Matthew Cleaveland et al.

Models of actual causality leverage domain knowledge to generate convincing diagnoses of events that caused an outcome. It is promising to apply these models to diagnose and repair run-time property violations in cyber-physical systems (CPS) with learning-enabled components (LEC). However, given the high diversity and complexity of LECs, it is challenging to encode domain knowledge (e.g., the CPS dynamics) in a scalable actual causality model that could generate useful repair suggestions. In this paper, we focus causal diagnosis on the input/output behaviors of LECs. Specifically, we aim to identify which subset of I/O behaviors of the LEC is an actual cause for a property violation. An important by-product is a counterfactual version of the LEC that repairs the run-time property by fixing the identified problematic behaviors. Based on this insights, we design a two-step diagnostic pipeline: (1) construct and Halpern-Pearl causality model that reflects the dependency of property outcome on the component's I/O behaviors, and (2) perform a search for an actual cause and corresponding repair on the model. We prove that our pipeline has the following guarantee: if an actual cause is found, the system is guaranteed to be repaired; otherwise, we have high probabilistic confidence that the LEC under analysis did not cause the property violation. We demonstrate that our approach successfully repairs learned controllers on a standard OpenAI Gym benchmark.

Kaustubh Sridhar, Souradeep Dutta, Ramneet Kaur et al.

Adversarial training (AT) and its variants have spearheaded progress in improving neural network robustness to adversarial perturbations and common corruptions in the last few years. Algorithm design of AT and its variants are focused on training models at a specified perturbation strength $ε$ and only using the feedback from the performance of that $ε$-robust model to improve the algorithm. In this work, we focus on models, trained on a spectrum of $ε$ values. We analyze three perspectives: model performance, intermediate feature precision and convolution filter sensitivity. In each, we identify alternative improvements to AT that otherwise wouldn't have been apparent at a single $ε$. Specifically, we find that for a PGD attack at some strength $δ$, there is an AT model at some slightly larger strength $ε$, but no greater, that generalizes best to it. Hence, we propose overdesigning for robustness where we suggest training models at an $ε$ just above $δ$. Second, we observe (across various $ε$ values) that robustness is highly sensitive to the precision of intermediate features and particularly those after the first and second layer. Thus, we propose adding a simple quantization to defenses that improves accuracy on seen and unseen adaptive attacks. Third, we analyze convolution filters of each layer of models at increasing $ε$ and notice that those of the first and second layer may be solely responsible for amplifying input perturbations. We present our findings and demonstrate our techniques through experiments with ResNet and WideResNet models on the CIFAR-10 and CIFAR-10-C datasets.

Mengyu Liu, Pengyuan Lu, Xin Chen et al.

We propose a model-free reinforcement learning solution, namely the ASAP-Phi framework, to encourage an agent to fulfill a formal specification ASAP. The framework leverages a piece-wise reward function that assigns quantitative semantic reward to traces not satisfying the specification, and a high constant reward to the remaining. Then, it trains an agent with an actor-critic-based algorithm, such as soft actor-critic (SAC), or deep deterministic policy gradient (DDPG). Moreover, we prove that ASAP-Phi produces policies that prioritize fulfilling a specification ASAP. Extensive experiments are run, including ablation studies, on state-of-the-art benchmarks. Results show that our framework succeeds in finding sufficiently fast trajectories for up to 97\% test cases and defeats baselines.

Xi Zheng, Aloysius K. Mok, Ruzica Piskac et al.

The integration of machine learning (ML) into cyber-physical systems (CPS) offers significant benefits, including enhanced efficiency, predictive capabilities, real-time responsiveness, and the enabling of autonomous operations. This convergence has accelerated the development and deployment of a range of real-world applications, such as autonomous vehicles, delivery drones, service robots, and telemedicine procedures. However, the software development life cycle (SDLC) for AI-infused CPS diverges significantly from traditional approaches, featuring data and learning as two critical components. Existing verification and validation techniques are often inadequate for these new paradigms. In this study, we pinpoint the main challenges in ensuring formal safety for learningenabled CPS.We begin by examining testing as the most pragmatic method for verification and validation, summarizing the current state-of-the-art methodologies. Recognizing the limitations in current testing approaches to provide formal safety guarantees, we propose a roadmap to transition from foundational probabilistic testing to a more rigorous approach capable of delivering formal assurance.

Wenwen Si, Sangdon Park, Insup Lee et al.

Prediction sets capture uncertainty by predicting sets of labels rather than individual labels, enabling downstream decisions to conservatively account for all plausible outcomes. Conformal inference algorithms construct prediction sets guaranteed to contain the true label with high probability. These guarantees fail to hold in the face of distribution shift, which is precisely when reliable uncertainty quantification can be most useful. We propose a novel algorithm for constructing prediction sets with PAC guarantees in the label shift setting. This method estimates the predicted probabilities of the classes in a target domain, as well as the confusion matrix, then propagates uncertainty in these estimates through a Gaussian elimination algorithm to compute confidence intervals for importance weights. Finally, it uses these intervals to construct prediction sets. We evaluate our approach on five datasets: the CIFAR-10, ChestX-Ray and Entity-13 image datasets, the tabular CDC Heart dataset, and the AGNews text dataset. Our algorithm satisfies the PAC guarantee while producing smaller, more informative, prediction sets compared to several baselines.

Seonwoo Kim, Jinwoo Kim, Daegyu Kang et al.

Cyber threat intelligence (CTI) reports now serve as essential resources for capturing adversary tactics, techniques, and procedures observed in modern attack campaigns. While traditional CTI platforms reduce this intelligence to isolated indicators through fixed schemas such as STIX, ontology-based representations preserve the semantic relationships needed for structured threat analysis. However, existing approaches for ontology-aligned CTI extraction face three challenges: (i) schema-specific pipelines that require manual reconfiguration whenever the schema changes, (ii) prompt-based schema inclusion that fails to scale on large ontologies such as UCO, and (iii) reliance on enterprise LLM APIs that conflicts with privacy constraints when integrating sensitive internal incident data. In this paper, we present ANCHOR, a schema-agnostic CTI knowledge graph construction system that bridges LLMs and formal ontology schemas. At the core of ANCHOR is hybrid ontology discovery, a search-and-navigate mechanism that dynamically explores large-scale ontology schemas, combined with SHACL-based validation to enforce schema-compliant type assignments. Experimental results on the UCO, STIX, and MALOnt schemas show that ANCHOR outperforms existing baselines in ontology typing and schema compliance. In addition, ANCHOR with a local LLM closely matches enterprise LLM typing performance, enabling privacy-preserving CTI analysis with high fidelity.

Kaustubh Sridhar, Souradeep Dutta, Dinesh Jayaraman et al.

Imitation learning considerably simplifies policy synthesis compared to alternative approaches by exploiting access to expert demonstrations. For such imitation policies, errors away from the training samples are particularly critical. Even rare slip-ups in the policy action outputs can compound quickly over time, since they lead to unfamiliar future states where the policy is still more likely to err, eventually causing task failures. We revisit simple supervised ``behavior cloning'' for conveniently training the policy from nothing more than pre-recorded demonstrations, but carefully design the model class to counter the compounding error phenomenon. Our ``memory-consistent neural network'' (MCNN) outputs are hard-constrained to stay within clearly specified permissible regions anchored to prototypical ``memory'' training samples. We provide a guaranteed upper bound for the sub-optimality gap induced by MCNN policies. Using MCNNs on 10 imitation learning tasks, with MLP, Transformer, and Diffusion backbones, spanning dexterous robotic manipulation and driving, proprioceptive inputs and visual inputs, and varying sizes and types of demonstration data, we find large and consistent gains in performance, validating that MCNNs are better-suited than vanilla deep neural networks for imitation learning applications. Website: https://sites.google.com/view/mcnn-imitation

Jean Park, Sydney Pugh, Kaustubh Sridhar et al.

Large Deep Neural Networks (DNNs) are often data hungry and need high-quality labeled data in copious amounts for learning to converge. This is a challenge in the field of medicine since high quality labeled data is often scarce. Data programming has been the ray of hope in this regard, since it allows us to label unlabeled data using multiple weak labeling functions. Such functions are often supplied by a domain expert. Data-programming can combine multiple weak labeling functions and suggest labels better than simple majority voting over the different functions. However, it is not straightforward to express such weak labeling functions, especially in high-dimensional settings such as images and time-series data. What we propose in this paper is a way to bypass this issue, using distance functions. In high-dimensional spaces, it is easier to find meaningful distance metrics which can generalize across different labeling tasks. We propose an algorithm that queries an expert for labels of a few representative samples of the dataset. These samples are carefully chosen by the algorithm to capture the distribution of the dataset. The labels assigned by the expert on the representative subset induce a labeling on the full dataset, thereby generating weak labels to be used in the data programming pipeline. In our medical time series case study, labeling a subset of 50 to 130 out of 3,265 samples showed 17-28% improvement in accuracy and 13-28% improvement in F1 over the baseline using clinician-defined labeling functions. In our medical image case study, labeling a subset of about 50 to 120 images from 6,293 unlabeled medical images using our approach showed significant improvement over the baseline method, Snuba, with an increase of approximately 5-15% in accuracy and 12-19% in F1 score.

Ramneet Kaur, Yiannis Kantaros, Wenwen Si et al.

Deep neural networks (DNN) have become a common sensing modality in autonomous systems as they allow for semantically perceiving the ambient environment given input images. Nevertheless, DNN models have proven to be vulnerable to adversarial digital and physical attacks. To mitigate this issue, several detection frameworks have been proposed to detect whether a single input image has been manipulated by adversarial digital noise or not. In our prior work, we proposed a real-time detector, called VisionGuard (VG), for adversarial physical attacks against single input images to DNN models. Building upon that work, we propose VisionGuard* (VG), which couples VG with majority-vote methods, to detect adversarial physical attacks in time-series image data, e.g., videos. This is motivated by autonomous systems applications where images are collected over time using onboard sensors for decision-making purposes. We emphasize that majority-vote mechanisms are quite common in autonomous system applications (among many other applications), as e.g., in autonomous driving stacks for object detection. In this paper, we investigate, both theoretically and experimentally, how this widely used mechanism can be leveraged to enhance the performance of adversarial detectors. We have evaluated VG* on videos of both clean and physically attacked traffic signs generated by a state-of-the-art robust physical attack. We provide extensive comparative experiments against detectors that have been designed originally for out-of-distribution data and digitally attacked images.

Yahan Yang, Soham Dan, Dan Roth et al.

Multilingual pre-trained Large Language Models (LLMs) are incredibly effective at Question Answering (QA), a core task in Natural Language Understanding, achieving high accuracies on several multilingual benchmarks. However, little is known about how well their confidences are calibrated. In this paper, we comprehensively benchmark the calibration of several multilingual LLMs (MLLMs) on a variety of QA tasks. We perform extensive experiments, spanning encoder-only, encoder-decoder, and decoder-only QA models (size varying from 110M to 7B parameters) and diverse languages, including both high- and low-resource ones. We study different dimensions of calibration in in-distribution, out-of-distribution, and cross-lingual transfer settings, and investigate strategies to improve it, including post-hoc methods and regularized fine-tuning. For decoder-only LLMs such as LlaMa2, we additionally find that in-context learning improves confidence calibration on multilingual data. We also conduct several ablation experiments to study the effect of language distances, language corpus size, and model size on calibration, and how multilingual models compare with their monolingual counterparts for diverse tasks and languages. Our experiments suggest that the multilingual QA models are poorly calibrated for languages other than English and incorporating a small set of cheaply translated multilingual samples during fine-tuning/calibration effectively enhances the calibration performance.

Xiayan Ji, Anton Xue, Eric Wong et al.

Anomaly detection is widely used for identifying critical errors and suspicious behaviors, but current methods lack interpretability. We leverage common properties of existing methods and recent advances in generative models to introduce counterfactual explanations for anomaly detection. Given an input, we generate its counterfactual as a diffusion-based repair that shows what a non-anomalous version should have looked like. A key advantage of this approach is that it enables a domain-independent formal specification of explainability desiderata, offering a unified framework for generating and evaluating explanations. We demonstrate the effectiveness of our anomaly explainability framework, AR-Pro, on vision (MVTec, VisA) and time-series (SWaT, WADI, HAI) anomaly datasets. The code used for the experiments is accessible at: https://github.com/xjiae/arpro.

Xinmeng Huang, Shuo Li, Mengxin Yu et al.

Language Models (LMs) have shown promising performance in natural language generation. However, as LMs often generate incorrect or hallucinated responses, it is crucial to correctly quantify their uncertainty in responding to given inputs. In addition to verbalized confidence elicited via prompting, many uncertainty measures ($e.g.$, semantic entropy and affinity-graph-based measures) have been proposed. However, these measures can differ greatly, and it is unclear how to compare them, partly because they take values over different ranges ($e.g.$, $[0,\infty)$ or $[0,1]$). In this work, we address this issue by developing a novel and practical framework, termed $Rank$-$Calibration$, to assess uncertainty and confidence measures for LMs. Our key tenet is that higher uncertainty (or lower confidence) should imply lower generation quality, on average. Rank-calibration quantifies deviations from this ideal relationship in a principled manner, without requiring ad hoc binary thresholding of the correctness score ($e.g.$, ROUGE or METEOR). The broad applicability and the granular interpretability of our methods are demonstrated empirically.

Kaustubh Sridhar, Souradeep Dutta, Dinesh Jayaraman et al.

Building generalist agents that can rapidly adapt to new environments is a key challenge for deploying AI in the digital and real worlds. Is scaling current agent architectures the most effective way to build generalist agents? We propose a novel approach to pre-train relatively small policies on relatively small datasets and adapt them to unseen environments via in-context learning, without any finetuning. Our key idea is that retrieval offers a powerful bias for fast adaptation. Indeed, we demonstrate that even a simple retrieval-based 1-nearest neighbor agent offers a surprisingly strong baseline for today's state-of-the-art generalist agents. From this starting point, we construct a semi-parametric agent, REGENT, that trains a transformer-based policy on sequences of queries and retrieved neighbors. REGENT can generalize to unseen robotics and game-playing environments via retrieval augmentation and in-context learning, achieving this with up to 3x fewer parameters and up to an order-of-magnitude fewer pre-training datapoints, significantly outperforming today's state-of-the-art generalist agents. Website: https://kaustubhsridhar.github.io/regent-research

Kaustubh Sridhar, Souradeep Dutta, Dinesh Jayaraman et al.

Multi-task ``vision-language-action'' (VLA) models have recently demonstrated increasing promise as generalist foundation models for robotics, achieving non-trivial performance out of the box on new tasks in new environments. However, for such models to be truly useful, an end user must have easy means to teach them to improve. For language and vision models, the emergent ability to perform in-context learning (ICL) has proven to be a versatile and highly useful interface to easily teach new tasks with no parameter finetuning. Unfortunately, VLAs pre-trained with imitation learning objectives do not naturally acquire ICL abilities. In this paper, we demonstrate that, with the right finetuning recipe and a small robot demonstration dataset, it is possible to inject in-context adaptability post hoc into such a VLA. After retraining for in-context learning (RICL), our system permits an end user to provide a small number (10-20) of demonstrations for a new task. RICL then fetches the most relevant portions of those demonstrations into the VLA context to exploit ICL, performing the new task and boosting task performance. We apply RICL to inject ICL into the $π_{0}$-FAST VLA, and show that it permits large in-context improvements for a variety of new manipulation tasks with only 20 demonstrations per task, without any parameter updates. When parameter updates on the target task demonstrations is possible, RICL finetuning further boosts performance. We release code and model weights for RICL-$π_{0}$-FAST alongside the paper to enable, for the first time, a simple in-context learning interface for new manipulation tasks. Website: https://ricl-vla.github.io.

Lianghuan Huang, Sagnik Anupam, Insup Lee et al.

Reinforcement learning (RL) has emerged as a promising strategy for finetuning small language models (SLMs) to solve targeted tasks such as math and coding. However, RL algorithms tend to be resource-intensive, taking a significant amount of time to train. We propose RAPID, a novel RL algorithm that can substantially reduce the running time of RL. Our key insight is that RL tends to be costly due to the need to perform both inference and backpropagation during training. To maximize use of computational resources, our algorithm performs inference in large batches, and then performs off-policy policy gradient updates in mini-batches. For off-policy updates, we incorporate group advantage estimation into the policy gradient algorithm, and derive an importance weighted estimator to correct for the bias arising from off-policy learning. Our experiments demonstrate that our algorithm can reduce running time by 11%-34% on three benchmarks compared to state-of-the-art RL algorithms while maintaining similar or better accuracy.

Sydney Pugh, Ivan Ruchkin, Insup Lee et al.

Deep learning models have shown promising predictive accuracy for time-series healthcare applications. However, ensuring the robustness of these models is vital for building trustworthy AI systems. 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, which is highly desirable for trustworthy AI. We propose a method to curate datasets comprised of natural adversarial examples to evaluate model robustness. The method relies on probabilistic labels obtained from automated weakly-supervised labeling that combines noisy and cheap-to-obtain labeling heuristics. Based on these labels, our method adversarially orders the input data and uses this ordering to construct a sequence of increasingly adversarial datasets. Our evaluation on six medical case studies and three non-medical case studies demonstrates the efficacy and statistical validity of our approach to generating naturally adversarial datasets

Souradeep Dutta, Kaustubh Sridhar, Osbert Bastani et al.

Long horizon robot learning tasks with sparse rewards pose a significant challenge for current reinforcement learning algorithms. A key feature enabling humans to learn challenging control tasks is that they often receive expert intervention that enables them to understand the high-level structure of the task before mastering low-level control actions. We propose a framework for leveraging expert intervention to solve long-horizon reinforcement learning tasks. We consider \emph{option templates}, which are specifications encoding a potential option that can be trained using reinforcement learning. We formulate expert intervention as allowing the agent to execute option templates before learning an implementation. This enables them to use an option, before committing costly resources to learning it. We evaluate our approach on three challenging reinforcement learning problems, showing that it outperforms state-of-the-art approaches by two orders of magnitude. Videos of trained agents and our code can be found at: https://sites.google.com/view/stickymittens

Ramneet Kaur, Susmit Jha, Anirban Roy et al.

Machine learning methods such as deep neural networks (DNNs), despite their success across different domains, are known to often generate incorrect predictions with high confidence on inputs outside their training distribution. The deployment of DNNs in safety-critical domains requires detection of out-of-distribution (OOD) data so that DNNs can abstain from making predictions on those. A number of methods have been recently developed for OOD detection, but there is still room for improvement. We propose the new method iDECODe, leveraging in-distribution equivariance for conformal OOD detection. It relies on a novel base non-conformity measure and a new aggregation method, used in the inductive conformal anomaly detection framework, thereby guaranteeing a bounded false detection rate. We demonstrate the efficacy of iDECODe by experiments on image and audio datasets, obtaining state-of-the-art results. We also show that iDECODe can detect adversarial examples.

Ivan Ruchkin, Matthew Cleaveland, Radoslav Ivanov et al.

Closed-loop verification of cyber-physical systems with neural network controllers offers strong safety guarantees under certain assumptions. It is, however, difficult to determine whether these guarantees apply at run time because verification assumptions may be violated. To predict safety violations in a verified system, we propose a three-step confidence composition (CoCo) framework for monitoring verification assumptions. First, we represent the sufficient condition for verified safety with a propositional logical formula over assumptions. Second, we build calibrated confidence monitors that evaluate the probability that each assumption holds. Third, we obtain the confidence in the verification guarantees by composing the assumption monitors using a composition function suitable for the logical formula. Our CoCo framework provides theoretical bounds on the calibration and conservatism of compositional monitors. Two case studies show that compositional monitors are calibrated better than their constituents and successfully predict safety violations.

Ramneet Kaur, Susmit Jha, Anirban Roy et al.

Deep neural networks (DNNs) are known to produce incorrect predictions with very high confidence on out-of-distribution inputs (OODs). This limitation is one of the key challenges in the adoption of DNNs in high-assurance systems such as autonomous driving, air traffic management, and medical diagnosis. This challenge has received significant attention recently, and several techniques have been developed to detect inputs where the model's prediction cannot be trusted. These techniques detect OODs as datapoints with either high epistemic uncertainty or high aleatoric uncertainty. We demonstrate the difference in the detection ability of these techniques and propose an ensemble approach for detection of OODs as datapoints with high uncertainty (epistemic or aleatoric). We perform experiments on vision datasets with multiple DNN architectures, achieving state-of-the-art results in most cases.

Sangdon Park, Edgar Dobriban, Insup Lee et al.

An important challenge facing modern machine learning is how to rigorously quantify the uncertainty of model predictions. Conveying uncertainty is especially important when there are changes to the underlying data distribution that might invalidate the predictive model. Yet, most existing uncertainty quantification algorithms break down in the presence of such shifts. We propose a novel approach that addresses this challenge by constructing \emph{probably approximately correct (PAC)} prediction sets in the presence of covariate shift. Our approach focuses on the setting where there is a covariate shift from the source distribution (where we have labeled training examples) to the target distribution (for which we want to quantify uncertainty). Our algorithm assumes given importance weights that encode how the probabilities of the training examples change under the covariate shift. In practice, importance weights typically need to be estimated; thus, we extend our algorithm to the setting where we are given confidence intervals for the importance weights. We demonstrate the effectiveness of our approach on covariate shifts based on DomainNet and ImageNet. Our algorithm satisfies the PAC constraint, and gives prediction sets with the smallest average normalized size among approaches that always satisfy the PAC constraint.

Kaustubh Sridhar, Oleg Sokolsky, Insup Lee et al.

Improving adversarial robustness of neural networks remains a major challenge. Fundamentally, training a neural network via gradient descent is a parameter estimation problem. In adaptive control, maintaining persistency of excitation (PoE) is integral to ensuring convergence of parameter estimates in dynamical systems to their true values. We show that parameter estimation with gradient descent can be modeled as a sampling of an adaptive linear time-varying continuous system. Leveraging this model, and with inspiration from Model-Reference Adaptive Control (MRAC), we prove a sufficient condition to constrain gradient descent updates to reference persistently excited trajectories converging to the true parameters. The sufficient condition is achieved when the learning rate is less than the inverse of the Lipschitz constant of the gradient of loss function. We provide an efficient technique for estimating the corresponding Lipschitz constant in practice using extreme value theory. Our experimental results in both standard and adversarial training illustrate that networks trained with the PoE-motivated learning rate schedule have similar clean accuracy but are significantly more robust to adversarial attacks than models trained using current state-of-the-art heuristics.

Taylor J. Carpenter, Radoslav Ivanov, Insup Lee et al.

This paper presents ModelGuard, a sampling-based approach to runtime model validation for Lipschitz-continuous models. Although techniques exist for the validation of many classes of models the majority of these methods cannot be applied to the whole of Lipschitz-continuous models, which includes neural network models. Additionally, existing techniques generally consider only white-box models. By taking a sampling-based approach, we can address black-box models, represented only by an input-output relationship and a Lipschitz constant. We show that by randomly sampling from a parameter space and evaluating the model, it is possible to guarantee the correctness of traces labeled consistent and provide a confidence on the correctness of traces labeled inconsistent. We evaluate the applicability and scalability of ModelGuard in three case studies, including a physical platform.

Sooyong Jang, Radoslav Ivanov, Insup Lee et al.

As machine learning techniques become widely adopted in new domains, especially in safety-critical systems such as autonomous vehicles, it is crucial to provide accurate output uncertainty estimation. As a result, many approaches have been proposed to calibrate neural networks to accurately estimate the likelihood of misclassification. However, while these methods achieve low calibration error, there is space for further improvement, especially in large-dimensional settings such as ImageNet. In this paper, we introduce a calibration algorithm, named Hoki, that works by applying random transformations to the neural network logits. We provide a sufficient condition for perfect calibration based on the number of label prediction changes observed after applying the transformations. We perform experiments on multiple datasets and show that the proposed approach generally outperforms state-of-the-art calibration algorithms across multiple datasets and models, especially on the challenging ImageNet dataset. Finally, Hoki is scalable as well, as it requires comparable execution time to that of temperature scaling.

Sooyong Jang, Insup Lee, James Weimer

Providing reliable model uncertainty estimates is imperative to enabling robust decision making by autonomous agents and humans alike. While recently there have been significant advances in confidence calibration for trained models, examples with poor calibration persist in most calibrated models. Consequently, multiple techniques have been proposed that leverage label-invariant transformations of the input (i.e., an input manifold) to improve worst-case confidence calibration. However, manifold-based confidence calibration techniques generally do not scale and/or require expensive retraining when applied to models with large input spaces (e.g., ImageNet). In this paper, we present the recursive lossy label-invariant calibration (ReCal) technique that leverages label-invariant transformations of the input that induce a loss of discriminatory information to recursively group (and calibrate) inputs - without requiring model retraining. We show that ReCal outperforms other calibration methods on multiple datasets, especially, on large-scale datasets such as ImageNet.

Sangdon Park, Shuo Li, Insup Lee et al.

A key challenge for deploying deep neural networks (DNNs) in safety critical settings is the need to provide rigorous ways to quantify their uncertainty. In this paper, we propose a novel algorithm for constructing predicted classification confidences for DNNs that comes with provable correctness guarantees. Our approach uses Clopper-Pearson confidence intervals for the Binomial distribution in conjunction with the histogram binning approach to calibrated prediction. In addition, we demonstrate how our predicted confidences can be used to enable downstream guarantees in two settings: (i) fast DNN inference, where we demonstrate how to compose a fast but inaccurate DNN with an accurate but slow DNN in a rigorous way to improve performance without sacrificing accuracy, and (ii) safe planning, where we guarantee safety when using a DNN to predict whether a given action is safe based on visual observations. In our experiments, we demonstrate that our approach can be used to provide guarantees for state-of-the-art DNNs.

Min Du, Nesime Tatbul, Brian Rivers et al.

Class distribution skews in imbalanced datasets may lead to models with prediction bias towards majority classes, making fair assessment of classifiers a challenging task. Metrics such as Balanced Accuracy are commonly used to evaluate a classifier's prediction performance under such scenarios. However, these metrics fall short when classes vary in importance. In this paper, we propose a simple and general-purpose evaluation framework for imbalanced data classification that is sensitive to arbitrary skews in class cardinalities and importances. Experiments with several state-of-the-art classifiers tested on real-world datasets from three different domains show the effectiveness of our framework - not only in evaluating and ranking classifiers, but also training them.

Sangdon Park, Osbert Bastani, James Weimer et al.

Reliable uncertainty estimates are an important tool for helping autonomous agents or human decision makers understand and leverage predictive models. However, existing approaches to estimating uncertainty largely ignore the possibility of covariate shift--i.e., where the real-world data distribution may differ from the training distribution. As a consequence, existing algorithms can overestimate certainty, possibly yielding a false sense of confidence in the predictive model. We propose an algorithm for calibrating predictions that accounts for the possibility of covariate shift, given labeled examples from the training distribution and unlabeled examples from the real-world distribution. Our algorithm uses importance weighting to correct for the shift from the training to the real-world distribution. However, importance weighting relies on the training and real-world distributions to be sufficiently close. Building on ideas from domain adaptation, we additionally learn a feature map that tries to equalize these two distributions. In an empirical evaluation, we show that our proposed approach outperforms existing approaches to calibrated prediction when there is covariate shift.

Yiannis Kantaros, Taylor Carpenter, Kaustubh Sridhar et al.

Deep neural network (DNN) models have proven to be vulnerable to adversarial digital and physical attacks. In this paper, we propose a novel attack- and dataset-agnostic and real-time detector for both types of adversarial inputs to DNN-based perception systems. In particular, the proposed detector relies on the observation that adversarial images are sensitive to certain label-invariant transformations. Specifically, to determine if an image has been adversarially manipulated, the proposed detector checks if the output of the target classifier on a given input image changes significantly after feeding it a transformed version of the image under investigation. Moreover, we show that the proposed detector is computationally-light both at runtime and design-time which makes it suitable for real-time applications that may also involve large-scale image domains. To highlight this, we demonstrate the efficiency of the proposed detector on ImageNet, a task that is computationally challenging for the majority of relevant defenses, and on physically attacked traffic signs that may be encountered in real-time autonomy applications. Finally, we propose the first adversarial dataset, called AdvNet that includes both clean and physical traffic sign images. Our extensive comparative experiments on the MNIST, CIFAR10, ImageNet, and AdvNet datasets show that VisionGuard outperforms existing defenses in terms of scalability and detection performance. We have also evaluated the proposed detector on field test data obtained on a moving vehicle equipped with a perception-based DNN being under attack.

Sangdon Park, Osbert Bastani, Nikolai Matni et al.

We propose an algorithm combining calibrated prediction and generalization bounds from learning theory to construct confidence sets for deep neural networks with PAC guarantees---i.e., the confidence set for a given input contains the true label with high probability. We demonstrate how our approach can be used to construct PAC confidence sets on ResNet for ImageNet, a visual object tracking model, and a dynamics model for the half-cheetah reinforcement learning problem.

Radoslav Ivanov, Taylor J. Carpenter, James Weimer et al.

This paper describes a verification case study on an autonomous racing car with a neural network (NN) controller. Although several verification approaches have been proposed over the last year, they have only been evaluated on low-dimensional systems or systems with constrained environments. To explore the limits of existing approaches, we present a challenging benchmark in which the NN takes raw LiDAR measurements as input and outputs steering for the car. We train a dozen NNs using two reinforcement learning algorithms and show that the state of the art in verification can handle systems with around 40 LiDAR rays, well short of a typical LiDAR scan with 1081 rays. Furthermore, we perform real experiments to investigate the benefits and limitations of verification with respect to the sim2real gap, i.e., the difference between a system's modeled and real performance. We identify cases, similar to the modeled environment, in which verification is strongly correlated with safe behavior. Finally, we illustrate LiDAR fault patterns that can be used to develop robust and safe reinforcement learning algorithms.

Mohammadhosein Hasanbeig, Yiannis Kantaros, Alessandro Abate et al.

Reinforcement Learning (RL) has emerged as an efficient method of choice for solving complex sequential decision making problems in automatic control, computer science, economics, and biology. In this paper we present a model-free RL algorithm to synthesize control policies that maximize the probability of satisfying high-level control objectives given as Linear Temporal Logic (LTL) formulas. Uncertainty is considered in the workspace properties, the structure of the workspace, and the agent actions, giving rise to a Probabilistically-Labeled Markov Decision Process (PL-MDP) with unknown graph structure and stochastic behaviour, which is even more general case than a fully unknown MDP. We first translate the LTL specification into a Limit Deterministic Buchi Automaton (LDBA), which is then used in an on-the-fly product with the PL-MDP. Thereafter, we define a synchronous reward function based on the acceptance condition of the LDBA. Finally, we show that the RL algorithm delivers a policy that maximizes the satisfaction probability asymptotically. We provide experimental results that showcase the efficiency of the proposed method.

Nima Roohi, Ramneet Kaur, James Weimer et al.

Industrial cyber-physical systems are hybrid systems with strict safety requirements. Despite not having a formal semantics, most of these systems are modeled using Stateflow/Simulink for mainly two reasons: (1) it is easier to model, test, and simulate using these tools, and (2) dynamics of these systems are not supported by most other tools. Furthermore, with the ever growing complexity of cyber-physical systems, grows the gap between what can be modeled using an automatic formal verification tool and models of industrial cyber-physical systems. In this paper, we present a simple formal model for self-deriving cars. While after some simplification, safety of this system has already been proven manually, to the best of our knowledge, no automatic formal verification tool supports its dynamics. We hope this serves as a challenge problem for formal verification tools targeting industrial applications.

Yazhou Tu, Zhiqiang Lin, Insup Lee et al.

Inertial sensors provide crucial feedback for control systems to determine motional status and make timely, automated decisions. Prior efforts tried to control the output of inertial sensors with acoustic signals. However, their approaches did not consider sample rate drifts in analog-to-digital converters as well as many other realistic factors. As a result, few attacks demonstrated effective control over inertial sensors embedded in real systems. This work studies the out-of-band signal injection methods to deliver adversarial control to embedded MEMS inertial sensors and evaluates consequent vulnerabilities exposed in control systems relying on them. Acoustic signals injected into inertial sensors are out-of-band analog signals. Consequently, slight sample rate drifts could be amplified and cause deviations in the frequency of digital signals. Such deviations result in fluctuating sensor output; nevertheless, we characterize two methods to control the output: digital amplitude adjusting and phase pacing. Based on our analysis, we devise non-invasive attacks to manipulate the sensor output as well as the derived inertial information to deceive control systems. We test 25 devices equipped with MEMS inertial sensors and find that 17 of them could be implicitly controlled by our attacks. Furthermore, we investigate the generalizability of our methods and show the possibility to manipulate the digital output through signals with relatively low frequencies in the sensing channel.

Sangdon Park, James Weimer, Insup Lee

Data-driven techniques are used in cyber-physical systems (CPS) for controlling autonomous vehicles, handling demand responses for energy management, and modeling human physiology for medical devices. These data-driven techniques extract models from training data, where their performance is often analyzed with respect to random errors in the training data. However, if the training data is maliciously altered by attackers, the effect of these attacks on the learning algorithms underpinning data-driven CPS have yet to be considered. In this paper, we analyze the resilience of classification algorithms to training data attacks. Specifically, a generic metric is proposed that is tailored to measure resilience of classification algorithms with respect to worst-case tampering of the training data. Using the metric, we show that traditional linear classification algorithms are resilient under restricted conditions. To overcome these limitations, we propose a linear classification algorithm with a majority constraint and prove that it is strictly more resilient than the traditional algorithms. Evaluations on both synthetic data and a real-world retrospective arrhythmia medical case-study show that the traditional algorithms are vulnerable to tampered training data, whereas the proposed algorithm is more resilient (as measured by worst-case tampering).