Detecting anomalous inputs, such as adversarial and out-of-distribution (OOD) inputs, is critical for classifiers deployed in real-world applications, especially deep neural network (DNN) classifiers that are known to be brittle on such inputs. We propose an unsupervised statistical testing framework for detecting such anomalous inputs to a trained DNN classifier based on its internal layer representations. By calculating test statistics at the input and intermediate-layer representations of the DNN, conditioned individually on the predicted class and on the true class of labeled training data, the method characterizes their class-conditional distributions on natural inputs. Given a test input, its extent of nonconformity with respect to the training distribution is captured using p-values of the class-conditional test statistics across the layers, which are then combined using a scoring function designed to score high on anomalous inputs. We focus on adversarial inputs, which are an important class of anomalous inputs, and also demonstrate the effectiveness of our method on general OOD inputs. The proposed framework also provides an alternative class prediction that can be used to correct the DNN's prediction on (detected) adversarial inputs. Experiments on well-known image classification datasets with strong adversarial attacks, including a custom attack method that uses the internal layer representations of the DNN, demonstrate that our method outperforms or performs comparably with five recently-proposed, competing detection methods.

Detecting out-of-distribution (OOD) inputs is critical for safely deploying deep learning models in an open-world setting. However, existing OOD detection solutions can be brittle in the open world, facing various types of adversarial OOD inputs. While methods leveraging auxiliary OOD data have emerged, our analysis reveals a key insight that the majority of auxiliary OOD examples may not meaningfully improve the decision boundary of the OOD detector. In this paper, we provide a theoretically motivated method, Adversarial Training with informative Outlier Mining (ATOM), which improves the robustness of OOD detection. We show that, by mining informative auxiliary OOD data, one can significantly improve OOD detection performance, and somewhat surprisingly, generalize to unseen adversarial attacks. ATOM achieves state-of-the-art performance under a broad family of natural and perturbed OOD evaluation tasks. For example, on the CIFAR-10 in-distribution dataset, ATOM reduces the FPR95 by up to 57.99% under adversarial OOD inputs, surpassing the previous best baseline by a large margin.

We consider representation learning (hypothesis class $\mathcal{H} = \mathcal{F}\circ\mathcal{G}$) where training and test distributions can be different. Recent studies provide hints and failure examples for domain invariant representation learning, a common approach for this problem, but the explanations provided are somewhat different and do not provide a unified picture. In this paper, we provide new decompositions of risk which give finer-grained explanations and clarify potential generalization issues. For Single-Source Domain Adaptation, we give an exact decomposition (an equality) of the target risk, via a natural hybrid argument, as sum of three factors: (1) source risk, (2) representation conditional label divergence, and (3) representation covariate shift. We derive a similar decomposition for the Multi-Source case. These decompositions reveal factors (2) and (3) as the precise reasons for failure to generalize. For example, we demonstrate that domain adversarial neural networks (DANN) attempt to regularize for (3) but miss (2), while a recent technique Invariant Risk Minimization (IRM) attempts to account for (2) but does not consider (3). We also verify our observations experimentally.

Machine learning (ML) algorithms, especially deep neural networks, have demonstrated success in several domains. However, several types of attacks have raised concerns about deploying ML in safety-critical domains, such as autonomous driving and security. An attacker perturbs a data point slightly in the concrete feature space (e.g., pixel space) and causes the ML algorithm to produce incorrect output (e.g. a perturbed stop sign is classified as a yield sign). These perturbed data points are called adversarial examples, and there are numerous algorithms in the literature for constructing adversarial examples and defending against them. In this paper we explore semantic adversarial examples (SAEs) where an attacker creates perturbations in the semantic space representing the environment that produces input for the ML model. For example, an attacker can change the background of the image to be cloudier to cause misclassification. We present an algorithm for constructing SAEs that uses recent advances in differential rendering and inverse graphics.

Manifold assumption in learning states that: the data lie approximately on a manifold of much lower dimension than the input space. Generative models learn to generate data according to the underlying data distribution. Generative models are used in various tasks, such as data augmentation and generating variation of images. This paper addresses the following question: do generative models need to be aware of the topology of the underlying data manifold in which the data lie? This paper suggests that the answer is yes and demonstrates that these can have ramifications on security-critical applications, such as generative-model based defenses for adversarial examples. We provide theoretical and experimental results to support our claims.

Directed graphical models (DGMs) are a class of probabilistic models that are widely used for predictive analysis in sensitive domains, such as medical diagnostics. In this paper we present an algorithm for differentially private learning of the parameters of a DGM with a publicly known graph structure over fully observed data. Our solution optimizes for the utility of inference queries over the DGM and \textit{adds noise that is customized to the properties of the private input dataset and the graph structure of the DGM}. To the best of our knowledge, this is the first explicit data-dependent privacy budget allocation algorithm for DGMs. We compare our algorithm with a standard data-independent approach over a diverse suite of DGM benchmarks and demonstrate that our solution requires a privacy budget that is $3\times$ smaller to obtain the same or higher utility.

Over recent years, devising classification algorithms that are robust to adversarial perturbations has emerged as a challenging problem. In particular, deep neural nets (DNNs) seem to be susceptible to small imperceptible changes over test instances. In this work, we study whether there is any learning task for which it is possible to design classifiers that are only robust against polynomial-time adversaries. Indeed, numerous cryptographic tasks (e.g. encryption of long messages) are only be secure against computationally bounded adversaries, and are indeed mpossible for computationally unbounded attackers. Thus, it is natural to ask if the same strategy could help robust learning. We show that computational limitation of attackers can indeed be useful in robust learning by demonstrating a classifier for a learning task in which computational and information theoretic adversaries of bounded perturbations have very different power. Namely, while computationally unbounded adversaries can attack successfully and find adversarial examples with small perturbation, polynomial time adversaries are unable to do so unless they can break standard cryptographic hardness assumptions. Our results, therefore, indicate that perhaps a similar approach to cryptography (relying on computational hardness) holds promise for achieving computationally robust machine learning. We also show that the existence of such learning task in which computational robustness beats information theoretic robustness implies (average case) hard problems in $\mathbf{NP}$.

Recent advances in Machine Learning (ML) have demonstrated that neural networks can exceed human performance in many tasks. While generalizing well over natural inputs, neural networks are vulnerable to adversarial inputs -an input that is ``similar'' to the original input, but misclassified by the model. Existing defenses focus on Lp-norm bounded adversaries that perturb ML inputs in the digital space. In the real world, however, attackers can generate adversarial perturbations that have a large Lp-norm in the digital space. Additionally, these defenses also come at a cost to accuracy, making their applicability questionable in the real world. To defend models against such a powerful adversary, we leverage one constraint on its power: the perturbation should not change the human's perception of the physical information; the physical world places some constraints on the space of possible attacks. Two questions follow: how to extract and model these constraints? and how to design a classification paradigm that leverages these constraints to improve robustness accuracy trade-off? We observe that an ML model is typically a part of a larger system with access to different input modalities. Utilizing these modalities, we introduce invariants that limit the attacker's action space. We design a hierarchical classification paradigm that enforces these invariants at inference time. As a case study, we implement and evaluate our proposal in the context of the real-world application of road sign classification because of its applicability to autonomous driving. With access to different input modalities, such as LiDAR, camera, and location we show how to extract invariants and develop a hierarchical classifier. Our results on the KITTI and GTSRB datasets show that we can improve the robustness against physical attacks at minimal harm to accuracy.

An emerging problem in trustworthy machine learning is to train models that produce robust interpretations for their predictions. We take a step towards solving this problem through the lens of axiomatic attribution of neural networks. Our theory is grounded in the recent work, Integrated Gradients (IG) [STY17], in axiomatically attributing a neural network's output change to its input change. We propose training objectives in classic robust optimization models to achieve robust IG attributions. Our objectives give principled generalizations of previous objectives designed for robust predictions, and they naturally degenerate to classic soft-margin training for one-layer neural networks. We also generalize previous theory and prove that the objectives for different robust optimization models are closely related. Experiments demonstrate the effectiveness of our method, and also point to intriguing problems which hint at the need for better optimization techniques or better neural network architectures for robust attribution training.

Attribution methods have been developed to explain the decision of a machine learning model on a given input. We use the Integrated Gradient method for finding attributions to define the causal neighborhood of an input by incrementally masking high attribution features. We study the robustness of machine learning models on benign and adversarial inputs in this neighborhood. Our study indicates that benign inputs are robust to the masking of high attribution features but adversarial inputs generated by the state-of-the-art adversarial attack methods such as DeepFool, FGSM, CW and PGD, are not robust to such masking. Further, our study demonstrates that this concentration of high-attribution features responsible for the incorrect decision is more pronounced in physically realizable adversarial examples. This difference in attribution of benign and adversarial inputs can be used to detect adversarial examples. Such a defense approach is independent of training data and attack method, and we demonstrate its effectiveness on digital and physically realizable perturbations.