Biblio

Many previous studies have shown that trust in automation mediates the effectiveness of automation in maintaining performance, and one critical factor that affects trust is the reliability of the automated system. In the cyber domain, automated systems are pervasive, yet the involvement of human trust has not been studied extensively as in other domains such as transportation.

In the current study, we used a phishing email identification task (with a phishing detection automated assistant system) as a testbed to study human trust in automation in the cyber domain. More specifically, we systematically investigated the influence of “description” (i.e., whether the user was informed about the actual reliability of the automated system) and “experience” (i.e., whether the user was provided feedback on their choices), in addition to the reliability level of the automated phishing detection system. These factors were varied in different conditions of response bias (false alarm vs. misses) and task difficulty (easy vs. difficult), which were found may be critical in a pilot study. Measures of user performance and trust were compared across different conditions. The measures of interest were human trust in the warning (a subjective rating of how trustable the warning system is), human reliance on the automated system (an objective measure of whether the participants comply with the system’s warnings), and performance (the overall quality of the decisions made).

Eight hundred eighty-seven phishing emails from Arizona State University, Brown University, and Cornell University were assessed by two reviewers for Cialdini’s six principles of persuasion: authority, social proof, liking/similarity, commitment/consistency, scarcity, and reciprocation. A correlational analysis of email characteristics by year revealed that the persuasion principles of commitment/consistency and scarcity have increased over time, while the principles of reciprocation and social proof have decreased over time. Authority and liking/similarity revealed mixed results with certain characteristics increasing and others decreasing. Results from this study can inform user training of phishing emails and help cybersecurity software to become more effective. 

The ineffectiveness of phishing warnings has been attributed to users' poor comprehension of the warning. However, the effectiveness of a phishing warning is typically evaluated at the time when users interact with a suspected phishing webpage, which we call the effect with phishing warning. Nevertheless, users' improved phishing detection when the warning is absent—or the effect of the warning—is the ultimate goal to prevent users from falling for phishing scams. We conducted an online study to evaluate the effect with and of several phishing warning variations, varying the point at which the warning was presented and whether procedural knowledge instruction was included in the warning interface. The current Chrome phishing warning was also included as a control. 360 Amazon Mechanical-Turk workers made submission; 500¬ word maximum for symposia) decisions about 10 login webpages (8 authentic, 2 fraudulent) with the aid of warning (first phase). After a short distracting task, the workers made the same decisions about 10 different login webpages (8 authentic, 2 fraudulent) without warning. In phase one, the compliance rates with two proposed warning interfaces (98% and 94%) were similar to those of the Chrome warning (98%), regardless of when the warning was presented. In phase two (without warning), performance was better for the condition in which warning with procedural knowledge instruction was presented before the phishing webpage in phase one, suggesting a better of effect than for the other conditions. With the procedural knowledge of how to determine a webpage’s legitimacy, users identified phishing webpages more accurately even without the warning being presented.

The ineffectiveness of phishing warnings has been attributed to users' poor comprehension of the warning. However, the effectiveness of a phishing warning is typically evaluated at the time when users interact with a suspected phishing webpage, which we call the effect with phishing warning. Nevertheless, users' improved phishing detection when the warning is absent—or the effect of the warning—is the ultimate goal to prevent users from falling for phishing scams. We conducted an online study to evaluate the effect with and of several phishing warning variations, varying the point at which the warning was presented and whether procedural knowledge instruction was included in the warning interface. The current Chrome phishing warning was also included as a control. 360 Amazon Mechanical-Turk workers made submission; 500¬ word maximum for symposia) decisions about 10 login webpages (8 authentic, 2 fraudulent) with the aid of warning (first phase). After a short distracting task, the workers made the same decisions about 10 different login webpages (8 authentic, 2 fraudulent) without warning. In phase one, the compliance rates with two proposed warning interfaces (98% and 94%) were similar to those of the Chrome warning (98%), regardless of when the warning was presented. In phase two (without warning), performance was better for the condition in which warning with procedural knowledge instruction was presented before the phishing webpage in phase one, suggesting a better of effect than for the other conditions. With the procedural knowledge of how to determine a webpage’s legitimacy, users identified phishing webpages more accurately even without the warning being presented.

To help establish a more scientific basis for security science, which will enable the development of fundamental theories and move the field from being primarily reactive to primarily proactive, it is important for research results to be reported in a scientifically rigorous manner. Such reporting will allow for the standard pillars of science, namely replication, meta-analysis, and theory building. In this paper we aim to establish a baseline of the state of scientific work in security through the analysis of indicators of scientific research as reported in the papers from the 2015 IEEE Symposium on Security and Privacy. To conduct this analysis, we developed a series of rubrics to determine the completeness of the papers relative to the type of evaluation used (e.g. case study, experiment, proof). Our findings showed that while papers are generally easy to read, they often do not explicitly document some key information like the research objectives, the process for choosing the cases to include in the studies, and the threats to validity. We hope that this initial analysis will serve as a baseline against which we can measure the advancement of the science of security.

Secure collaboration requires the collaborating parties to apply the
right policies for their interaction.  We adopt a notion of
conditional, directed norms as a way to capture the standards of
correctness for a collaboration.  How can we handle conflicting norms?
We describe an approach based on knowledge of what norm dominates what
norm in what situation.  Our approach adapts answer-set programming to
compute stable sets of norms with respect to their computed conflicts
and dominance.  It assesses agent compliance with respect to those
stable sets.  We demonstrate our approach on a healthcare scenario.

Norms provide a way to model the social architecture of a sociotechnical system (STS) and are thus crucial for understanding how such a system supports secure collaboration between principals,that is, autonomous parties such as humans and organizations. Accordingly, an important challenge is to compute the state of a norm instance at runtime in a sociotechnical system.

Custard addresses this challenge by providing a relational syntax for schemas of important norm types along with their canonical lifecycles and providing a mapping from each schema to queries that compute instances of the schema in different lifecycle stages.  In essence, Custard supports a norm-based abstraction layer over underlying information stores such as databases and event logs. Specifically, it supports deadlines; complex events, including those based on aggregation; and norms that reference other norms.

We prove important correctness properties for Custard, including stability (once an event has occurred, it has occurred forever) and safety (a query returns a finite set of tuples).  Our compiler generates SQL queries from Custard specifications.  Writing out such SQL queries by hand is tedious and error-prone even for simple norms, thus demonstrating Custard's practical benefits.

Firewall policies are notorious for having misconfiguration errors which can defeat its intended purpose of protecting hosts in the network from malicious users. We believe this is because today's firewall policies are mostly monolithic. Inspired by ideas from modular programming and code refactoring, in this work we introduce three kinds of modules: primary, auxiliary, and template, which facilitate the refactoring of a firewall policy into smaller, reusable, comprehensible, and more manageable components. We present algorithms for generating each of the three modules for a given legacy firewall policy. We also develop ModFP, an automated tool for converting legacy firewall policies represented in access control list to their modularized format. With the help of ModFP, when examining several real-world policies with sizes ranging from dozens to hundreds of rules, we were able to identify subtle errors.

Firewall policies are notorious for having misconfiguration errors which can defeat its intended purpose of protecting hosts in the network from malicious users. We believe this is because today's firewall policies are mostly monolithic. Inspired by ideas from modular programming and code refactoring, in this work we introduce three kinds of modules: primary, auxiliary, and template, which facilitate the refactoring of a firewall policy into smaller, reusable, comprehensible, and more manageable components. We present algorithms for generating each of the three modules for a given legacy firewall policy. We also develop ModFP, an automated tool for converting legacy firewall policies represented in access control list to their modularized format. With the help of ModFP, when examining several real-world policies with sizes ranging from dozens to hundreds of rules, we were able to identify subtle errors.

To interact effectively, agents must enter into commitments. What should an agent do when these commitments conflict? We describe Coco, an approach for reasoning about which specific commitments apply to specific parties in light of general types of commitments, specific circumstances, and dominance relations among specific commitments. Coco adapts answer-set programming to identify a maximalsetofnondominatedcommitments. It provides a modeling language and tool geared to support practical applications.

In a multiagent system, a (social) norm describes what the agents may expect from each other.  Norms promote autonomy (an agent need not comply with a norm) and heterogeneity (a norm describes interactions at a high level independent of implementation details). Researchers have studied norm emergence through social learning where the agents interact repeatedly in a graph structure.

In contrast, we consider norm emergence in an open system, where membership can change, and where no predetermined graph structure exists.  We propose Silk, a mechanism wherein a generator monitors interactions among member agents and recommends norms to help resolve conflicts.  Each member decides on whether to accept or reject a recommended norm.  Upon exiting the system, a member passes its experience along to incoming members of the same type.  Thus, members develop norms in a hybrid manner to resolve conflicts.

We evaluate Silk via simulation in the traffic domain.  Our results show that social norms promoting conflict resolution emerge in both moderate and selfish societies via our hybrid mechanism.

Smartphone users often use private and enterprise data with untrusted third party applications.  The fundamental lack of secrecy guarantees in smartphone OSes, such as Android, exposes this data to the risk of unauthorized exfiltration.  A natural solution is the integration of secrecy guarantees into the OS.  In this paper, we describe the challenges for decentralized information flow control (DIFC) enforcement on Android.  We propose context-sensitive DIFC enforcement via lazy polyinstantiation and practical and secure network export through domain declassification.  Our DIFC system, Weir, is backwards compatible by design, and incurs less than 4 ms overhead for component startup.  With Weir,  we demonstrate practical and secure DIFC enforcement on Android.

Platform as a Service (PaaS) provides middleware resources to cloud customers. As demand for PaaS services increases, so do concerns about the security of PaaS. This paper discusses principal PaaS security and integrity requirements, and vulnerabilities and the corresponding countermeasures. We consider three core cloud elements: multi-tenancy, isolation, and virtualization and how they relate to PaaS services and security trends and concerns such as user and resource isolation, side-channel vulnerabilities in multi-tenant environments, and protection of sensitive data

Signals intelligence analysts play a critical role in the United States government by providing information regarding potential national security threats to government leaders. Analysts perform complex decision-making tasks that involve gathering, sorting, and analyzing information. The current study evaluated how individual differences and training influence performance on an Internet search-based medical diagnosis task designed to simulate a signals analyst task. The implemented training emphasized the extraction and organization of relevant information and deductive reasoning. The individual differences of interest included working memory capacity and previous experience with elements of the task, specifically health literacy, prior experience using the Internet, and prior experience conducting Internet searches. Preliminary results indicated that the implemented training did not significantly affect performance, however, working memory significantly predicted performance on the implemented task. These results support previous research and provide additional evidence that working memory capacity influences performance on cognitively complex decision-making tasks, whereas experience with elements of the task may not. These findings suggest that working memory capacity should be considered when screening individuals for signals intelligence positions. Future research should aim to generalize these findings within a broader sample, and ideally utilize a task that directly replicates those performed by signals analysts.

An Assessment of Security Problems in Open Source Software: Improving software security through changes in software design and development processes appears to be a very hard problem. For example, well documented security issues such as Structured Query Language injection, after more than a decade, still tops most vulnerability lists. Security priority is often subdued due to constraints such as time-to-market and resources. Furthermore, security process outcomes are hard to quantify and even harder to predict or relate to process improvement activities. In part this is because of the nature of the security faults - they are in statistical terms "rare" and often very complex compared to "regular" non-security faults. In part it is the irregular and unpredictable nature of the security threats and attacks that puts the software under attack into states it was not designed for and subjects it to what would be considered "nonoperational" use. In many cases it is the human component of the system that fails - for example, due to phishing or due to incorrect use of a software product. On the other hand, we have decades of experience developing reliable software (admittedly subject to similar resource, cost and time constraints). The central question of interest in this thesis is to what extent can we leverage some of the software reliability engineering (SRE) models, processes, and metrics that work in the "classical" operational space to develop predictive software security engineering assessment and development elements. Specific objectives are a) to investigate use of (possibly modified) SRE practices to characterize security properties of software, and b) assess how software design and development processes could be enhanced to avoid, eliminate and tolerate security problems and attacks.We are particularly interested in open source software security, the conditions under which SRE practices may be useful, and the information that this can provide about the security quality of a software product. We examined public information about security problem reports for open source Fedora and RHEL series of software releases, Chromium project and Android project. The data that we analyzed was primarily about security problems reported from post-release in-the-field use of the products. What can we learn about the non-operational processes (and possible threats) related to security problems? One aspect is classification of security problems based on the traits that contribute to the injection of problems into code, whether due to poor practices or limited knowledge (epistemic errors), or due to random accidental events (aleatoric errors). Knowing the distribution can help understand attack space and help improve development processes and testing of the next version. For example, in the case of Fedora, the distribution of security problems found post-release was consistent across two different releases of the software. The security problem discovery rate appears to be roughly constant but much lower (ten to a hundred times lower) than the initial non-security problem discovery rate. Similarly, in the case of RHEL, the distribution of security problems found post-release was consistent and the number of security problems kept decreasing across six different releases of the software. The security problem discovery rate appears to be roughly constant but again much lower than the initial non-security problem discovery rate. In the case of Chromium, the number of discovered security problems is orders of magnitude higher than for other products, except that does not appear to translate into a higher incidence of field breaches. One reason could be Chromium "bounty" for problem discovery. We find that some classical reliability models can be used as one of tools to estimate the residual number of security problems in both the current release and in the future releases of the software, and through that provide a measure of the security characteristics of the software. For example, to assess whether, under given usage conditions, security problem discovery rate is increasing or decreasing - and what that may mean. Based on our findings, we discuss an agile software testing process that combines operational and non-operational (or attack related) testing with the intent of finding and eliminating more security problems earlier in the software development process. The knowledge of vulnerable components from architectural view and the frequency of vulnerabilities in each of the components helps in prioritizing security test resources.

We propose Cupid, a language for specifying commitments that supports their information-centric aspects, and offers crucial benefits.  One, Cupid is first-order, enabling a systematic treatment of commitment instances.  Two, Cupid supports features needed for real-world scenarios such as deadlines, nested commitments, and complex event expressions for capturing the lifecycle of commitment instances.  Three, Cupid maps to relational database queries and thus provides a set-based semantics for retrieving commitment instances in states such as being violated,discharged, and so on.  We prove that Cupid queries are safe.  Four,to aid commitment modelers, we propose the notion of well-identified commitments, and finitely violable and finitely expirable commitments.  We give syntactic restrictions for obtaining such commitments.

Automated cyber attacks tend to be schedule and resource limited. The primary progress metric is often “coverage” of pre-determined “known” vulnerabilities that may not have been patched, along with possible zero-day exploits (if such exist). We present and discuss a hypergeometric process model that describes such attack patterns. We used web request signatures from the logs of a production web server to assess the applicability of the model.

Phishing is a social engineering tactic that targets internet users in an attempt to trick them into divulging personal information. When opening an email, users are faced with the decision of determining if an email is legitimate or an attempt at phishing. Although software has been developed to assist the user, studies have shown they are not foolproof, leaving the user vulnerable. Multiple training programs have been developed to educate users in their efforts to make informed decisions; however, training that conveys the real world consequences of phishing or training that increases a user’s fear level have not been developed. Conveying real world consequences of a situation and increasing a user’s fear level have been proven to enhance the effects of training in other fields. Ninety-six participants were recruited and randomly assigned to training programs with phishing consequences, training programs designed to increase fear, or a control group. Preliminary results indicate that training helped users identify phishing emails; however, little difference was seen among the three groups. Future analysis will include a factor analysis of personality and individual differences that influence training efficacy.

State estimation plays a critically important role in ensuring the secure and reliable operation of the electric grid. Recent works have shown that the state estimation process is vulnerable to stealthy attacks where an adversary can alter certain measurements to corrupt the solution of the process, but evade the existing bad data detection algorithms and remain invisible to the system operator. Since the state estimation result is used to compute optimal power flow and perform contingency analysis, incorrect estimation can undermine economic and secure system operation. However, an adversary needs sufficient resources as well as necessary knowledge to achieve a desired attack outcome. The knowledge that is required to launch an attack mainly includes the measurements considered in state estimation, the connectivity among the buses, and the power line admittances. Uncertainty in information limits the potential attack space for an attacker. This advantage of uncertainty enables us to apply moving target defense (MTD) strategies for developing a proactive defense mechanism for state estimation.

In this paper, we propose an MTD mechanism for securing state estimation, which has several characteristics: (i) increase the knowledge uncertainty for attackers, (ii) reduce the window of attack opportunity, and (iii) increase the attack cost. In this mechanism, we apply controlled randomization on the power grid system properties, mainly on the set of measurements that are considered in state estimation, and the topology, especially the line admittances. We thoroughly analyze the performance of the proposed mechanism on the standard IEEE 14- and 30-bus test systems.

Growing traffic volumes and the increasing complexity of attacks pose a constant scaling challenge for network intrusion prevention systems (NIPS). In this respect, offloading NIPS processing to compute clusters offers an immediately deployable alternative to expensive hardware upgrades. In practice, however, NIPS offloading is challenging on three fronts in contrast to passive network security functions: (1) NIPS offloading can impact other traffic engineering objectives; (2) NIPS offloading impacts user perceived latency; and (3) NIPS actively change traffic volumes by dropping unwanted traffic. To address these challenges, we present the SNIPS system. We design a formal optimization framework that captures tradeoffs across scalability, network load, and latency. We provide a practical implementation using recent advances in software-defined networking without requiring modifications to NIPS hardware. Our evaluations on realistic topologies show that SNIPS can reduce the maximum load by up to 10× while only increasing the latency by 2%.

In our previous work we showed that for Fedora, under normal operational conditions, security problem discovery appears to be a random process. While in the case of Fedora, and a number of other open source products, classical reliability models can be adapted to estimate the number of residual security problems under “normal” operational usage (not attacks), the predictive ability of the model is lower for security faults due to the rarity of security events and because there appears to be no real security reliability growth. The ratio of security to non-security faults is an indicator that the process needs improving, but it also may be leveraged to assess vulnerability profile of a release and possibly guide testing of its next version. We manually analyzed randomly sampled problems for four different versions of Fedora and classified them into security vulnerability categories. We also analyzed the distribution of these problems over the software’s lifespan and we found that they exhibit a symmetry which can perhaps be used in estimating the number of residual security problems in the software. Based on our work, we believe that an approach to vulnerability elimination based on a combination of “classical” and some non-operational “bounded” high-assurance testing along the lines discussed in may yield good vulnerability elimination results, as well as a way of estimating vulnerability level of a release. Classical SRE methods, metrics and models can be used to track both non-security and security problem detection under normal operational profile. We can then model the reliability growth, if any, and estimate the number of residual faults by estimating the lower and upper bounds on the total number of faults of a certain type. In production, there may be a simpler alternative. Just count the vulnerabilities and project over the next period assuming constant vulnerability discovery rate. In testing phase, to accelerate the process, one might leverage collected vulnerability information to generate non-operational test-cases aimed at vulnerability categories. The observed distributions of security problems reported under normal “operational” usage appear to support the above approach – i.e., what is learned say in the first x weeks can them be leveraged in selecting test cases in the next stage. Similarly, what is learned about a product Y weeks after its release may be very indicative of its vulnerability profile for the rest of its life given the assumption of constant vulnerability discovery rate.

Software as a Service (SaaS) is the most prevalent service delivery mode for cloud systems. This paper surveys common security vulnerabilities and corresponding countermeasures for SaaS. It is primarily focused on the work published in the last five years. We observe current SaaS security trends and a lack of sufficiently broad and robust countermeasures in some of the SaaS security area such as Identity and Access management due to the growth of SaaS applications.