Countering the substantial threat posed by antimicrobial resistance (AMR) requires the capability to predict pathogen resistance phenotype quickly and accurately. Deriving computational models to form such predictions is challenging for several reasons, perhaps the most daunting being the limited number of reliably-labeled examples from which to learn. We overcome this obstacle with a novel lightly-supervised algorithm that leverages unlabeled and/or noisily-labeled examples - which are often more readily-available - to facilitate model induction. Importantly, we prove the algorithm is safe: adding unlabeled/unreliably-labeled data to the learning procedure produces models which are usually more accurate, and never less accurate, than models learned with reliably-labeled data alone. Experiments demonstrate that the proposed method outperforms state-of-the-art models for the task of predicting AMR using pathogen genome sequences.

This paper describes a method for inferring when a person might be coordinating with others based on their behavior. We show that, in Wikipedia, editing behavior is more random when coordinating with others. We analyzed this using both entropy and conditional entropy. These algorithms rely only on timestamped events associated with entities, making them broadly applicable to other domains. In this paper, we will discuss previous research on this topic, how we adapted that research to the problem ofWikipedia edit behavior, describe how we extended it, and discuss our results.

Machine learning (ML) plays a central role in the solution of many security problems, for example enabling malicious and innocent activities to be rapidly and accurately distinguished and appropriate actions to be taken. Unfortunately, a standard assumption in ML - that the training and test data are identically distributed - is typically violated in security applications, leading to degraded algorithm performance and reduced security. Previous research has attempted to address this challenge by developing ML algorithms which are either robust to differences between training and test data or are able to predict and account for these differences. This paper adopts a different approach, developing a class of moving target (MT) defenses that are difficult for adversaries to reverse-engineer, which in turn decreases the adversaries' ability to generate training/test data differences that benefit them. We leverage the coevolutionary relationship between attackers and defenders to derive a simple, flexible MT defense strategy which is optimal or nearly optimal for a broad range of security problems. Case studies involving two distinct cyber defense applications demonstrate that the proposed MT algorithm outperforms standard static methods, offering effective defense against intelligent, adaptive adversaries.

Many important security informatics problems require consideration of dynamical phenomena for their solution; examples include predicting the behavior of individuals in social networks and distinguishing malicious and innocent computer network activities based on activity traces. While information theory offers powerful tools for analyzing dynamical processes, to date the application of information-theoretic methods in security domains has focused on static analyses (e.g., cryptography, natural language processing). This paper leverages information-theoretic concepts and measures to quantify the similarity of pairs of stochastic dynamical systems, and shows that this capability can be used to solve important problems which arise in security applications. We begin by presenting a concise review of the information theory required for our development, and then address two challenging tasks: 1.) characterizing the way influence propagates through social networks, and 2.) distinguishing malware from legitimate software based on the instruction sequences of the disassembled programs. In each application, case studies involving real-world datasets demonstrate that the proposed techniques outperform standard methods.

Previous attempts at supply chain risk management are often non-technical and rely heavily on policies/procedures to provide security assurances. This is particularity worrisome as there are vast volumes of data that must be analyzed and data continues to grow at unprecedented rates. In order to mitigate these issues and minimize the amount of manual inspection required, we propose the development of mathematically-based automated screening methods that can be incorporated into supply chain risk management. In particular, we look at methods for identifying deception and deceptive practices that may be present in the supply chain. We examine two classes of constraints faced by deceivers, cognitive/computational limitations and strategic tradeoffs, which can be used to developed graph-based metrics to represent entity behavior. By using these metrics with novel machine learning algorithms, we can robustly detect deceptive behavior and identify potential supply chain issues.

Adaptive adversaries are a primary concern in several domains, including cyber defense, border security, counterterrorism, and fraud prevention, and consequently there is great interest in developing defenses that maintain their effectiveness in the presence of evolving adversary strategies and tactics. This paper leverages the coevolutionary relationship between attackers and defenders to derive two new approaches to predictive defense, in which future attack techniques are anticipated and these insights are incorporated into defense designs. The first method combines game theory with machine learning to model and predict future adversary actions in the learner's “feature space”; these predictions form the basis for synthesizing robust defenses. The second approach to predictive defense involves extrapolating the evolution of defense configurations forward in time, in the space of defense parameterizations, as a way of generating defenses which work well against evolving threats. Case studies with a large cyber security dataset assembled for this investigation demonstrate that each method provides effective, scalable defense against current and future attacks, outperforming gold-standard techniques. Additionally, preliminary tests indicate that a simple variant of the proposed design methodology yields defenses which are difficult for adversaries to reverse-engineer.

There is considerable interest in developing techniques for predicting human behavior, for instance to enable emerging contentious situations to be anticipated or permit the nature of ongoing but “hidden” activities to be inferred. A promising approach to this problem is to collect appropriate empirical data and then apply machine learning methods to the data to generate the predictions. This two-part paper shows that the performance of such learning algorithms often can be improved substantially by leveraging sociological models in their development and implementation. In particular, we demonstrate that sociologically-grounded learning algorithms outperform gold-standard methods in two important and challenging tasks: 1.) inferring the (unobserved) nature of relationships in adversarial social networks, and 2.) predicting whether nascent social diffusion events will “go viral”. Significantly, the new algorithms perform well even when there is limited data available for their training and execution.

There is considerable interest in developing techniques for predicting human behavior, and a promising approach to this problem is to collect phenomenon-relevant empirical data and then apply machine learning methods to these data to form predictions. This two-part paper shows that the performance of such learning algorithms often can be improved substantially by leveraging sociological models in their development and implementation. In this paper, the second of the two parts, we demonstrate that a sociologically-grounded learning algorithm outperforms a gold-standard method for the task of predicting whether nascent social diffusion events will “go viral”. Significantly, the proposed algorithm performs well even when there is only limited time series data available for analysis.

A frame in social theory is a schema of interpretation that individuals rely upon to understand and respond to events. Collective action frames are persuasive communications constructed during mobilization campaigns, frequently to call attention to a perceived problem or issue, itemize steps to resolve the issue, and motivate others regarding the issue's urgency and the importance of taking action [e.g. 1]. There is substantial empirical evidence that framing rhetoric influences the success or failure of social movements, and that movement members are increasingly relying upon social media to communicate their frames. Consequently, there is considerable interest within the security informatics community to develop methods for detecting framing rhetoric in social (and other) media [2]. Particularly valuable would be computational methods capable of distinguishing framing from non-framing content using only limited labeled examples of framing discourse, because generating such labels typically requires experienced analysts and is an expensive and time-consuming undertaking.

There are substantial potential benefits to considering predictability when designing defenses against adaptive adversaries, including increasing the ability of defense systems to predict new attacker behavior and reducing the capacity of adversaries to anticipate defensive actions. This paper adopts such a perspective, leveraging the coevolutionary relationship between attackers and defenders to derive methods for predicting and countering attacks and for limiting the extent to which adversaries can learn about defense strategies. The proposed approach combines game theory with machine learning to model adversary adaptation in the learner's feature space, thereby producing classes of predictive and “moving target” defenses which are scientifically-grounded and applicable to problems of real-world scale and complexity. Case studies with large cyber security datasets demonstrate that the proposed algorithms outperform gold-standard techniques, offering effective and robust defense against evolving adversaries.

There is significant interest to develop proactive approaches to cyber defense, in which future attack strategies are anticipated and these insights are incorporated into defense designs. This paper considers the problem of protecting computer networks against intrusions and other attacks, and leverages the coevolutionary relationship between attackers and defenders to derive two new methods for proactive network defense. The first method is a bipartite graph-based machine learning algorithm which enables information concerning previous attacks to be “transferred” for application against novel attacks, thereby substantially increasing the rate with which defense systems can successfully respond to new attacks. The second approach involves exploiting basic threat information (e.g., from cyber security analysts) to generate “synthetic” attack data for use in training defense systems, resulting in networks defenses that are effective against both current and (near) future attacks. The utility of the proposed methods is demonstrated by showing that they outperform standard techniques for the task of detecting malicious network activity in two publicly-available cyber datasets.

Discovering and characterizing emerging topics and trends through analysis of Web data is of great interest to security analysts and policy makers. This paper considers the problem of monitoring social media to spot emerging memes - distinctive phrases which act as “tracers” for discrete cultural units - as a means of rapidly detecting new topics and trends. We have recently developed a method for predicting which memes will propagate widely and which will not, thereby enabling the discovery of significant topics. Here we demonstrate the efficacy of this approach through case studies involving political memes and memes associated with an emerging cyber threat.

Inferring the sentiment of social media content, for instance blog posts and forum threads, is both of great interest to security analysts and technically challenging to accomplish. This paper presents two computational methods for estimating social media sentiment which address the challenges associated with Web-based analysis. Each method formulates the task as one of text classification, models the data as a bipartite graph of documents and words, and assumes that only limited prior information is available regarding the sentiment orientation of any of the documents or words of interest. The first algorithm is a semi-supervised sentiment classifier which combines knowledge of the sentiment labels for a few documents and words with information present in unlabeled data, which is abundant online. The second algorithm assumes existence of a set of labeled documents in a domain related to the domain of interest, and leverages these data to estimate sentiment in the target domain. We demonstrate the utility of the proposed methods by showing they outperform several standard techniques for the task of inferring the sentiment of online movie and consumer product reviews. Additionally, we illustrate the potential of the methods for security informatics by estimating regional public opinion regarding Egypt's unfolding revolution through analysis of Arabic, Indonesian, and Danish (language) blog posts.

There is considerable interest in developing predictive capabilities for social diffusion processes, for instance enabling early identification of contentious “triggering” incidents that are likely to grow into large, self-sustaining mobilization events. Recently we have shown, using theoretical analysis, that the dynamics of social diffusion may depend crucially upon the interactions of social network communities, that is, densely connected groupings of individuals which have only relatively few links to other groups. This paper presents an empirical investigation of two hypotheses which follow from this finding: 1.) the presence of even just a few inter-community links can make diffusion activity in one community a significant predictor of activity in otherwise disparate communities and 2.) very early dispersion of a diffusion process across network communities is a reliable early indicator that the diffusion will ultimately involve a substantial number of individuals. We explore these hypotheses with case studies involving emergence of the Swedish Social Democratic Party at the turn of the 20th century, the spread of SARS in 2002-2003, and blogging dynamics associated with potentially incendiary real world occurrences. These empirical studies demonstrate that network community-based diffusion metrics do indeed possess predictive power, and in fact can be significantly more predictive than standard measures.

This paper presents a computational approach to inferring the sentiment orientation of “social media” content (e.g., blog posts) which focuses on the challenges associated with Web-based analysis. The proposed methodology formulates the task as one of text classification, models the data as a bipartite graph of documents and words, and uses this framework to develop a semi-supervised sentiment classifier that is well-suited for social media domains. In particular, the proposed algorithm is capable of combining prior knowledge regarding the sentiment orientation of a few documents and words with information present in unlabeled data, which is abundant online. We demonstrate the utility of the approach by showing it outperforms several standard methods for the task of inferring the sentiment of online movie reviews, and illustrate its potential for security informatics through a case study involving the estimation of Indonesian public sentiment regarding the July 2009 Jakarta hotel bombings.

The Internet occasionally experiences large disruptions, arising from both natural and manmade disturbances, and it is of significant interest to develop methods for locating within the network the source of a given disruption (i.e., the network element(s) whose perturbation initiated the event). This paper presents a near real-time approach to realizing this logical localization objective. The proposed methodology consists of three steps: 1.) data acquisition/preprocessing, in which publicly available measurements of Internet activity are acquired, “cleaned”, and assembled into a format suitable for computational analysis, 2.) event characterization via tensor factorization-based time series analysis, and 3.) localization of the source of the disruption through graph theoretic analysis. This procedure provides a principled, automated approach to identifying the root causes of network disruptions at “whole-Internet” scale. The considerable potential of the proposed analytic method is illustrated through a computer simulation study and empirical analysis of a recent, large-scale Internet disruption.

This paper considers the position regulation problem for uncertain robot manipulators in the presence of constraints on the available actuator torques, and proposes two new controllers as solutions to this problem. The first controller is derived under the assumption that the manipulator state is measurable, while the second strategy is developed for those applications in which only position measurements are available. Each scheme consists of a nonadaptive component for gross position control and an adaptive component to ensure convergence to the desired position. The controllers are computationally simple, require very little information regarding the manipulator model or the payload, and ensure that the position error is globally convergent. The capabilities of the proposed control strategies are illustrated though both computer simulations and laboratory experiments with an IMI Zebra Zero manipulator.

This paper considers the problem of controlling the motion of manipulators to perform waste management tasks. In particular, we develop two real-time control modules for implementation with the Sequential Modular Architecture for Robotics and Teleoperation (SMART) which has been developed at Sandia National Laboratories. SMART provides a "generic" framework within which individual subsystems can be combined, simulated, and implemented for real-time control of complex robotic systems. Control modules developed in the paper include an adaptive gravity compensation controller and a real-time sensor-based collision avoidance strategy. The adaptive gravity compensation controller is a position regulation strategy that requires no information regarding the manipulator dynamic model or payload and can be proven to be passive. The whole-arm collision avoidance algorithm is implemented within a damped-least-squares inverse kinematic formulation. The example modules are demonstrated through computer simulation on a 15 degree-of-freedom manipulator.

This paper presents a new adaptive controller as a solution to the problem of stabilizing nonholonomic mechanical systems in the presence of incomplete information concerning the system dynamic model. The proposed control system consists of two subsystems: a slightly modified version of the kinematic stabilization strategy of M'Closkey and Murray, which generates a desired velocity trajectory for the nonholonomic system, and an adaptive control scheme which ensures that this velocity trajectory is accurately tracked. This approach is shown to provide arbitrarily accurate stabilization to any desired configuration and can be implemented with no knowledge of the system dynamic model. The efficacy of the proposed stabilization strategy is illustrated through computer simulations with two nonholonomic mechanical systems.

This paper considers the problem of controlling the motion of highly redundant manipulators to achieve a desired set of tasks. The paper develops two real-time control modules for implementation with the Sequential Modular Architecture for Robotics and Teleoperation (SMART) which has been developed at Sandia National Laboratories. SMART provides a "generic" framework within which individual subsystems can be combined, simulated, and implemented for real-time control of complex robotic systems. Control modules developed in the paper include an adaptive gravity compensation controller and a real-time redundancy resolution strategy. The adaptive gravity compensation controller is a position regulation strategy that requires no information regarding the manipulator dynamic model or payload and can be proven to be passive. The redundancy resolution strategy uses a damped-least-squares (DLS) inverse kinematic formulation. The example controller modules are demonstrated through computer simulation on a 20-DOF serpentine manipulator.

This paper considers the position regulation problem for uncertain robot manipulators and proposes two adaptive controllers as solutions to this problem. The first controller is derived under the assumption that the manipulator state is measurable, while the second strategy is developed for those applications in which only position measurements are available. Each of the adaptive schemes is computationally simple, requires very little information regarding the manipulator model or the payload, and ensures that the position error possesses desirable convergence properties: semiglobal asymptotic convergence if no external disturbances are present and semiglobal convergence to an arbitrarily small neighbourhood of zero in the presence of bounded disturbances. Experimental results are presented for a Zebra Zero manipulator and demonstrate that accurate and robust position control can be achieved by using the proposed controllers.

This paper presents two adaptive schemes for compliant motion control of uncertain manipulators: an adaptive impedance controller and an adaptive position/force controller. The proposed controllers are very general and computationally efficient, need virtually no information regarding the manipulator dynamic model or the environment, and require only position and force measurements for implementation. It is shown that the schemes ensure semiglobal uniform boundedness of all signals in the presence of bounded disturbances, and that the ultimate size of the system errors can be made arbitrarily small. The capabilities of the proposed control strategies are illustrated though both computer simulations and laboratory experiments with a six degree-of-freedom (DOF) IMI Zebra Zero manipulator.

Presents a robust strategy for estimating tool wear in metal cutting operations. The proposed estimation algorithm consists of two components: a recurrent neural network to model the tool wear dynamics, and a robust observer to estimate the tool wear from this model using measurements of cutting force. It is shown that the algorithm ensures that the tool wear estimation error is uniformly bounded in the presence of bounded unmodeled effects, and that the ultimate bound on this error can be made as small as desired. The proposed approach is applied to the problem of estimating tool wear in turning and is shown to provide wear estimates which are in close agreement with published experimental results.

This paper considers the position regulation problem for uncertain rigid-link electrically-driven manipulators, and presents an adaptive control strategy as a solution to this problem. The proposed controller is computationally simple, requires virtually no a priori information concerning either the manipulator or actuator models, and ensures semiglobal stability and convergence of the manipulator position error to zero. Experimental results are presented for a Zebra Zero manipulator and demonstrate that the proposed scheme provides a simple and effective means of obtaining high performance position regulation. Additionally, the experiments indicate that accurate trajectory tracking can also be realized with this controller.

This paper presents a simple and robust approach to achieving collision avoidance for kinematically redundant manipulators at the control-loop level. The proposed scheme represents the obstacle avoidance requirement as inequality constraints in the manipulator workspace, and ensures that these inequalities are satisfied while the end-effector tracks the desired trajectory. The control scheme is the damped-least-squares formulation of the configuration control approach implemented as a kinematic controller. Computer simulation and experimental results are given for a Robotics Research 7 DOF redundant arm and demonstrate the collision avoidance capability for reaching inside a truss structure. These results confirm that the proposed approach provides a simple and effective method for real-time collision avoidance.>

This paper presents two adaptive schemes for compliance control of redundant manipulators. The first scheme is developed using an adaptive impedance control approach for torque-controlled manipulators, while the second strategy is an adaptive admittance controller for position-controlled manipulators. The proposed controllers are very general and computationally efficient since they do not require knowledge of the manipulator dynamic model or parameter values of the manipulator or the environment, and are implemented without calculation of the inverse dynamics or inverse kinematic transformation. It is shown that the control strategies are globally stable in the presence of bounded disturbances, and that in the absence of disturbances the ultimate bound on the size of the system errors can be made arbitrarily small. The capabilities of the proposed admittance controller are illustrated through laboratory experiments with a Robotics Research Corporation seven degree-of-freedom (DOF) manipulator.>