Deepali Arora

STARS: A Framework for Statistically Rigorous Simulation-Based Network Research

Simulation has become one of the dominant tools in wired and wireless network research. With the advent of cloud, grid, and cluster computing it has become feasible to use parallelization to perform richer larger-scale simulations. Moreover, the computing resources needed to perform statistically rigorous simulations are now easily obtainable. Although a number of parallel network simulation frameworks exists, the issue of statistical rigorous testing has largely not been addressed. This work presents a parallel MPI-aware network simulation framework that is specifically designed to provide automated support for statistically rigorous experimentation, thereby offloading this significant researcher burden. Unlike prior frameworks, the proposed framework includes a distribution-free statistical analysis feedback loop that automatically deduces the next set of experiments that need to be run. The value of this new framework is highlighted by exploring the well known issue of assessing the true duration of start-up transients within mobile ad hoc networks (MANETs) simulations.

Analytics: Key to Go from Generating Big Data to Deriving Business Value

The potential to extract actionable insights from Big Data has gained increased attention of researchers in academia as well as several industrial sectors. The field has become interesting and problems look even more exciting to solve ever since organizations have been trying to tame large volumes of complex and fast arriving Big Data streams through newer computing paradigms. However, extracting meaningful and actionable information from Big Data is a challenging and daunting task. The ability to generate value from large volumes of data is an art which combined with analytical skills needs to be mastered in order to gain competitive advantage in business. The ability of organizations to leverage the emerging technologies and integrate Big Data into their enterprise architectures effectively depends on the maturity level of the technology and business teams, capabilities they develop as well as the strategies they adopt. In this paper, through selected use cases, we demonstrate how statistical analyses, machine learning algorithms, optimization and text mining algorithms can be applied to extract meaningful insights from the data available through social media, online commerce, telecommunication industry, smart utility meters and used for variety of business benefits, including improving security. The nature of applied analytical techniques largely depends on the underlying nature of the problem so a one-size-fits-all solution hardly exists. Deriving information from Big Data is also subject to challenges associated with data security and privacy. These and other challenges are discussed in context of the selected problems to illustrate the potential of Big Data analytics.

Assessing the Expected Performance of the OLSR Routing Protocol for Denser Urban Core Ad Hoc Network Deployments

The wide-scale adoption of smart phones has begun to provide a pragmatic real-world deployment environments for mobile ad hoc networks, (i.e., as peer-to-peer game platforms, for emergency services, etc.). Such deployments are likely to occur with urban cores where device densities would easily exceed those that have traditionally been studied. Moreover, the quality of the resulting solutions will innately rest on the capabilities of the underlying routing protocols. Of current protocols, the OLSR proactive routing protocol makes the strongest arguments regarding its suitability to such larger, denser network environments. This work tests OLSRs true suitability by analyzing its performance within a 360-node network existing within a standard 1 km ×1.5 km communications area, (i.e., innately for a network with approximately 3 × the node densities typically studied). It is shown that OLSR largely fails for such denser networks, with these failure arising due to OLSRs underlying presumption that routing tables updates should occur relatively infrequently. This limitation within OLSR has not been previously reported and this work highlights the reasons why these issues were likely not observed within prior OLSR studies.

On the statistical behaviors of network-level features within MANETs

Event-based simulation has become a primary means of pursuing mobile ad hoc network (MANET) research. The stochastic nature of MANETs has been well studied with respect to mobility models, but less work has looked at the statistical behaviors of network layer features, (e.g., PDR, delay, hops and routing overhead). Fundamentally, issues such as “When do start up transients end?” and “Do all Monte- Carlo runs indeed arrive at the same steady-state distributions?” have not been well explored. This work explores these issues through using the DYMO routing protocol and the OMNeT++ simulation framework as exemplars. By applying distribution free Kolmogorov-Smirnov goodness-of-fit tests it is shown that, for network-layer features: a) MANET start-up transients can persist far longer than previously reported, b) transient durations can vary significantly from feature to feature and with varying node velocities, and c) Monte-Carlo runs of a given MANET scenario can produce distinct behavioral modes. It is then discussed whether these issues are likely inherent to MANETs and their routing protocols or an artifact of OMNeT++.

Big Data Analytics for Classification of Network Enabled Devices

As information technology (IT) and telecommunication systems continue to grow in size and complexity, especially with Internet of Things (IoT) gaining popularity, maintaining a secure and seamless exchange of information between devices becomes a challenging task. A large number of devices connected over the Internet leads to an increase in vulnerabilities and security threats, which makes the identification of critical assets necessary. Asset identification helps organizations to identify and to respond quickly to any security breaches. In this paper, machine learning based techniques are used to identify assets based on their connectivity, i.e., servers and endpoints. For the analysis presented in this paper four different machine learning algorithms, K-Nearest Neighbor, Naive Bayes, Support Vector Machines, and Random Forest algorithms are used and the performance of these algorithms is assessed in terms of the F-score calculated for each algorithm. Results show that for a given dataset, amongst all four algorithms, the Random Forest classifier achieved highest accuracy in terms of identifying the assets correctly. However, the Random Forest algorithm is computationally intensive and may not work for large datasets. Naive Bayes algorithm yielded the worst performance and KNearest Neighbors performance was very close to that achieved by Support Vector Machines. Our results shows that for the given dataset, Support Vector Machine based classifier was found to be a good compromise in terms of accuracy and computational expensiveness.

Mining WiFi Data for Business Intelligence

The WiFi networks provide an ease of accessing email, Web, and other Internet applications while on the move. However, deploying additional WiFi hotspots that can provide both increased coverage and enhance user quality of service largely depends upon the number of access points already existing and user densities. Extracting usage patterns and information from the available data has the potential to answer several business-focussed questions. In this paper, we show that by plotting WiFi locations in a two-dimensional space of incoming (downloading) and outgoing (uploading) data amount, in conjunction with the simple k-means clustering, it is possible to gain insight into the basic data usage patterns. When combined with information about geographic location of the WiFi hotspots such analysis can answer questions related to spatial patterns of data usage and make informed business decisions including charging customers at selected locations for WiFi service.

Assessing Trade-Offs between Stealthiness and Node Recruitment Rates in Peer-to-Peer Botnets

Botnets denote collections of compromised computers under adversary control and, although early botnets using centralized command and control (C&C) structures were fairly easily defeated, botnets remain a serious global security threat. in part, this is due to the evolution within the adversarial communities using highly diffuse decentralized peer-to-peer (P2P) based C&C within modern botnets, which has proven far more difficult to address. the resulting increased botnet resilience though comes at the cost of placing the bots further from the botmasterâs direct control, thereby, increasing the time required to recruit subsets of bots to specific malicious tasks, (i.e., to send spam, engage in a DDOS attack, etc.). This work explores the specific tradeoffs that occur between achievable bot recruitment rates and overall botnet stealthiness within P2P structured botnets. It is shown that rapid recruitment of nodes (or bots) leads directly to an order of magnitude increase in the botnetâs generated network traffic, which makes the botnet significantly more visible (and susceptible) to defensive counter-measures. Kademlia is used through out this work as the exemplar P2P protocol as, within the real-world, Kademlia has proven to provide an effective C&C mechanism for a number of the longer-lived botnets.

Lower bounds on mobile terminal localisation in an urban area

Lower bounds on localisation errors serve as a performance indicator of how close a localisation system is to providing optimal performance. In this study, localisation of mobile terminals for urban areas is performed using received signal strength (RSS)-based techniques with maximum likelihood (ML) and linear kernel (LK) estimators. Simulations are performed with and without buildings in an urban area cell to illustrate the effect of discontinuities in the RSS profiles, on radio location accuracy. Results show that a localisation error is higher when buildings are absent as compared to the scenario when buildings are present. Buildings add extra features to the RSS measurement space which, if known to the localisation system, improve radio location accuracy. A comparison is made between the root mean-square error of the ML and LK estimators with the Cramer–Rao bound (CRB), the Bayesian Cramer–Rao bound (BCRB) and the Weiss–Weinstein bound (WWB). These comparisons show that the previously used CRB and BCRB do not provide realistic lowers bounds in the presence of buildings. In such cases bounds, such as WWB, which are capable of handling RSS discontinuities provide more realistic lower bounds on the accuracy of radio location.

Statistical Assessment of Sybil-Placement Strategies within DHT-Structured Peer-to-Peer Botnets

Botnets are a well recognized global cyber-security threat as they enable attack communities to command large collections of compromised computers (bots) on-demand. Peer to-peer (P2P) distributed hash tables (DHT) have become particularly attractive botnet command and control (C & C) solutions due to the high level resiliency gained via the diffused random graph overlays they produce. The injection of Sybils, computers pretending to be valid bots, remains a key defensive strategy against DHT-structured P2P botnets. This research uses packet level network simulations to explore the relative merits of random, informed, and partially informed Sybil placement strategies. It is shown that random placements perform nearly as effectively as the tested more informed strategies, which require higher levels of inter-defender co-ordination. Moreover, it is shown that aspects of the DHT-structured P2P botnets behave as statistically nonergodic processes, when viewed from the perspective of stochastic processes. This suggests that although optimal Sybil placement strategies appear to exist they would need carefully tuning to each specific P2P botnet instance.

Enabling richer statistical MANET simulations through cluster computing

The wide-scale adoption of modern smart phones and other multi-radio mobile devices, has begun to provide pragmatic deployment environments for non-cellular mobile ad hoc network (MANET) services (i.e., for disaster recovery scenarios, peered mobile games, social networking applications, etc.). User perceptions of the quality of such MANET services will be driven, in part, by standard network-level quality of service (QoS) metrics such as delay, jitter, throughput, etc. Much of the existing MANET literature has explored these issues, as well as MANET routing protocol design, through single computer Monte Carlo simulations (e.g., via ns-2, ns-3, OMNeT++, or OpNet). Results are then reported as the averages of these Monte Carlo runs. As is well known from probability and statistics, such averaging is only meaningful when applied across statistically ergodic data (i.e., data drawn from the same underlying distribution). But, assessing the validity of this underlying ergodic assumption requires transitioning to more rigorous cluster-based MANET simulation frameworks. This work highlights the theoretical rationale for such ergodicity testing, the developments of a cluster-based framework, the STARs framework, to support such testing, and the results and insights obtained by using this framework to evaluate the popular DYMO and OLSR MANET routing protocols. This work also discusses why the insights ergodic testing provides are of interest to potential real-world MANET deployments.