Recent advances in machine learning, especially techniques such as deep neural networks, are promoting a range of high-stakes applications, including autonomous driving, which often relies on deep learning for perception. While deep learning for perception has been shown to be vulnerable to a host of subtle adversarial manipulations of images, end-to-end demonstrations of successful attacks, which manipulate the physical environment and result in physical consequences, are scarce. Moreover, attacks typically involve carefully constructed adversarial examples at the level of pixels. We demonstrate the first end-to-end attacks on autonomous driving in simulation, using simple physically realizable attacks: the painting of black lines on the road. These attacks target deep neural network models for end-to-end autonomous driving control. A systematic investigation shows that such attacks are surprisingly easy to engineer, and we describe scenarios (e.g., right turns) in which they are highly effective, and others that are less vulnerable (e.g., driving straight). Further, we use network deconvolution to demonstrate that the attacks succeed by inducing activation patterns similar to entirely different scenarios used in training.
The problem of dispatching emergency responders to service traffic accidents, fire, distress calls and crimes plagues urban areas across the globe. While such problems have been extensively looked at, most approaches are offline. Such methodologies fail to capture the dynamically changing environments under which critical emergency response occurs, and therefore, fail to be implemented in practice. Any holistic approach towards creating a pipeline for effective emergency response must also look at other challenges that it subsumes - predicting when and where incidents happen and understanding the changing environmental dynamics. We describe a system that collectively deals with all these problems in an online manner, meaning that the models get updated with streaming data sources. We highlight why such an approach is crucial to the effectiveness of emergency response, and present an algorithmic framework that can compute promising actions for a given decision-theoretic model for responder dispatch. We argue that carefully crafted heuristic measures can balance the trade-off between computational time and the quality of solutions achieved and highlight why such an approach is more scalable and tractable than traditional approaches. We also present an online mechanism for incident prediction, as well as an approach based on recurrent neural networks for learning and predicting environmental features that affect responder dispatch. We compare our methodology with prior state-of-the-art and existing dispatch strategies in the field, which show that our approach results in a reduction in response time with a drastic reduction in computational time.
An important problem in networked systems is detection and removal of suspected malicious nodes. A crucial consideration in such settings is the uncertainty endemic in detection, coupled with considerations of network connectivity, which impose indirect costs from mistakely removing benign nodes as well as failing to remove malicious nodes. A recent approach proposed to address this problem directly tackles these considerations, but has a significant limitation: it assumes that the decision maker has accurate knowledge of the joint maliciousness probability of the nodes on the network. This is clearly not the case in practice, where such a distribution is at best an estimate from limited evidence. To address this problem, we propose a distributionally robust framework for optimal node removal. While the problem is NP-Hard, we propose a principled algorithmic technique for solving it approximately based on duality combined with Semidefinite Programming relaxation. A combination of both theoretical and empirical analysis, the latter using both synthetic and real data, provide strong evidence that our algorithmic approach is highly effective and, in particular, is significantly more robust than the state of the art.
A fundamental challenge in networked systems is detection and removal of suspected malicious nodes. In reality, detection is always imperfect, and the decision about which potentially malicious nodes to remove must trade off false positives (erroneously removing benign nodes) and false negatives (mistakenly failing to remove malicious nodes). However, in network settings this conventional tradeoff must now account for node connectivity. In particular, malicious nodes may exert malicious influence, so that mistakenly leaving some of these in the network may cause damage to spread. On the other hand, removing benign nodes causes direct harm to these, and indirect harm to their benign neighbors who would wish to communicate with them. We formalize the problem of removing potentially malicious nodes from a network under uncertainty through an objective that takes connectivity into account. We show that optimally solving the resulting problem is NP-Hard. We then propose a tractable solution approach based on a convex relaxation of the objective. Finally, we experimentally demonstrate that our approach significantly outperforms both a simple baseline that ignores network structure, as well as a state-of-the-art approach for a related problem, on both synthetic and real-world datasets.
Despite the considerable success of convolutional neural networks in a broad array of domains, recent research has shown these to be vulnerable to small adversarial perturbations, commonly known as adversarial examples. Moreover, such examples have shown to be remarkably portable, or transferable, from one model to another, enabling highly successful black-box attacks. We explore this issue of transferability and robustness from two dimensions: first, considering the impact of conventional $l_p$ regularization as well as replacing the top layer with a linear support vector machine (SVM), and second, the value of combining regularized models into an ensemble. We show that models trained with different regularizers present barriers to transferability, as does partial information about the models comprising the ensemble.
Machine learning (ML) techniques are increasingly common in security applications, such as malware and intrusion detection. However, there is increasing evidence that machine learning models are susceptible to evasion attacks, in which an adversary makes small changes to the input (such as malware) in order to cause erroneous predictions (for example, to avoid being detected). Evasion attacks on ML fall into two broad categories: 1) those which generate actual malicious instances and demonstrate both evasion of ML and efficacy of attack (we call these problem space attacks), and 2) attacks which directly manipulate features used by ML, abstracting efficacy of attack into a mathematical cost function (we call these feature space attacks). Central to our inquiry is the following fundamental question: are feature space models of attacks useful proxies for real attacks? In the process of answering this question, we make two major contributions: 1) a general methodology for evaluating validity of mathematical models of ML evasion attacks, and 2) an application of this methodology as a systematic hypothesis-driven evaluation of feature space evasion attacks on ML-based PDF malware detectors. Specific to our case study, we find that a) feature space evasion models are in general not adequate in representing real attacks, b) such models can be significantly improved by identifying conserved features (features that are invariant in real attacks) whenever these exist, and c) ML hardened using the improved feature space models remains robust to alternative attacks, in contrast to ML hardened using a very powerful class of problem space attacks, which does not.
Despite the considerable success enjoyed by machine learning techniques in practice, numerous studies demonstrated that many approaches are vulnerable to attacks. An important class of such attacks involves adversaries changing features at test time to cause incorrect predictions. Previous investigations of this problem pit a single learner against an adversary. However, in many situations an adversary's decision is aimed at a collection of learners, rather than specifically targeted at each independently. We study the problem of adversarial linear regression with multiple learners. We approximate the resulting game by exhibiting an upper bound on learner loss functions, and show that the resulting game has a unique symmetric equilibrium. We present an algorithm for computing this equilibrium, and show through extensive experiments that equilibrium models are significantly more robust than conventional regularized linear regression.
We develop a novel optical neural network (ONN) framework which introduces a degree of scalar invariance to image classification estima- tion. Taking a hint from the human eye, which has higher resolution near the center of the retina, images are broken out into multiple levels of varying zoom based on a focal point. Each level is passed through an identical convolutional neural network (CNN) in a Siamese fashion, and the results are recombined to produce a high accuracy estimate of the object class. ONNs act as a wrapper around existing CNNs, and can thus be applied to many existing algorithms to produce notable accuracy improvements without having to change the underlying architecture.
Attacks in cyber-physical systems (CPS) which manipulate sensor readings can cause enormous physical damage if undetected. Detection of attacks on sensors is crucial to mitigate this issue. We study supervised regression as a means to detect anomalous sensor readings, where each sensor's measurement is predicted as a function of other sensors. We show that several common learning approaches in this context are still vulnerable to \emph{stealthy attacks}, which carefully modify readings of compromised sensors to cause desired damage while remaining undetected. Next, we model the interaction between the CPS defender and attacker as a Stackelberg game in which the defender chooses detection thresholds, while the attacker deploys a stealthy attack in response. We present a heuristic algorithm for finding an approximately optimal threshold for the defender in this game, and show that it increases system resilience to attacks without significantly increasing the false alarm rate.
For enhancing the privacy protections of databases, where the increasing amount of detailed personal data is stored and processed, multiple mechanisms have been developed, such as audit logging and alert triggers, which notify administrators about suspicious activities; however, the two main limitations in common are: 1) the volume of such alerts is often substantially greater than the capabilities of resource-constrained organizations, and 2) strategic attackers may disguise their actions or carefully choosing which records they touch, making incompetent the statistical detection models. For solving them, we introduce a novel approach to database auditing that explicitly accounts for adversarial behavior by 1) prioritizing the order in which types of alerts are investigated and 2) providing an upper bound on how much resource to allocate for each type. We model the interaction between a database auditor and potential attackers as a Stackelberg game in which the auditor chooses an auditing policy and attackers choose which records to target. A corresponding approach combining linear programming, column generation, and heuristic search is proposed to derive an auditing policy. For testing the policy-searching performance, a publicly available credit card application dataset are adopted, on which it shows that our methods produce high-quality mixed strategies as database audit policies, and our general approach significantly outperforms non-game-theoretic baselines.