Misty Blowers

Machine Learning Applied to Cyber Operations

Cyber attacks have evolved from operational to strategic events, with the aim to disrupt and influence strategic capability and assets, impede business operations, and target physical assets and mission critical information. With this emerging sophistication, current Intrusion Detection Systems (IDS) are also constantly evolving. As new viruses have emerged, the technologies used to detect them have also become more complex relying on sophisticated heuristics. Hosts and networks are constantly evolving with both security upgrades and topology changes. In addition, at most critical points of vulnerability, there are often vigilant humans in the loop.

On detection and visualization techniques for cyber security situation awareness

Networking technologies are exponentially increasing to meet worldwide communication requirements. The rapid growth of network technologies and perversity of communications pose serious security issues. In this paper, we aim to developing an integrated network defense system with situation awareness capabilities to present the useful information for human analysts. In particular, we implement a prototypical system that includes both the distributed passive and active network sensors and traffic visualization features, such as 1D, 2D and 3D based network traffic displays. To effectively detect attacks, we also implement algorithms to transform real-world data of IP addresses into images and study the pattern of attacks and use both the discrete wavelet transform (DWT) based scheme and the statistical based scheme to detect attacks. Through an extensive simulation study, our data validate the effectiveness of our implemented defense system.

An open-set speaker identification system using genetic learning classifier system

This paper presents the design and implementation of an adaptive open-set speaker identification system with genetic learning classifier systems. One of the challenging problems in using learning classifier systems for numerical problems is the knowledge representation. The voice samples are a series of real numbers that must be encoded in a classifier format. We investigate several different methods for representing voice samples for classifier systems and study the efficacy of the methods. We also identify several challenges for learning classifier systems in the speaker identification problem and introduce new methods to improve the learning and classification abilities of the systems. Experimental results show that our system successfully learns 200 voice features at the accuracies of 60% to 80%, which is considered a strong result in the speaker identification community. This research presents the feasibility of using learning classifier systems for the speaker identification problem.

Multivariable analysis, correlation, and prediction

Making best use of multi-point observations and sensor information to forecast future events in complex real time systems is a challenge which presents itself in many military and industrial problem domains. The first step in tackling these challenges is to analyze and understand the data. Depending on the algorithm used to forecast a future event, improvements to a prediction can be realized if one can first determine the nature and extent of variable correlations, and for the purposes of prediction, quantify the strength of the correlations of input variables to output variables. This is no easy task since sensor readings and operator logs are sometimes inconsistent and/or unreliable, some catastrophic failures can be almost impossible to predict, and time lags and leads in a given system may vary from one day to the next. Correlation analysis techniques can help us deal with some of these problems. They allow us to find out what variables may be strongly correlated to major events. After detecting where the strongest correlations exist, one must choose a model which can best predict the possible outcomes that could occur for a number of possible scenarios. The model must be tested and evaluated, and sometimes it is necessary to go back to the feature selection stage of the model design process and reevaluate the available sensory data and inputs. An industrial process example is adopted in this research to both highlight the issues that arise in complex systems and to demonstrate methods of addressing such issues.

Toward Open-Set Text-Independent Speaker Identification in Tactical Communications

We present the design and implementation of an open-set text-independent speaker identification system using genetic learning classifier systems (LCS). We examine the use of this system in a real-number problem domain, where there is strong interest in its application to tactical communications. We investigate different encoding methods for representing real-number knowledge and study the efficacy of each method for speaker identification. We also identify several difficulties in solving the speaker identification problems with LCS and introduce new approaches to resolve the difficulties. Experimental results show that our system successfully learns 200 voice features at accuracies of 90 % to 100 % and 15,000 features to more than 80% for the closed-set problem, which is considered a strong result in the speaker identification community. The open-set capability is also comparable to existing numeric-based methods

Evaluating data distribution and drift vulnerabilities of machine learning algorithms in secure and adversarial environments

Machine learning is continuing to gain popularity due to its ability to solve problems that are difficult to model using conventional computer programming logic. Much of the current and past work has focused on algorithm development, data processing, and optimization. Lately, a subset of research has emerged which explores issues related to security. This research is gaining traction as systems employing these methods are being applied to both secure and adversarial environments. One of machine learning’s biggest benefits, its data-driven versus logic-driven approach, is also a weakness if the data on which the models rely are corrupted. Adversaries could maliciously influence systems which address drift and data distribution changes using re-training and online learning. Our work is focused on exploring the resilience of various machine learning algorithms to these data-driven attacks. In this paper, we present our initial findings using Monte Carlo simulations, and statistical analysis, to explore the maximal achievable shift to a classification model, as well as the required amount of control over the data.

Open-set speaker identification with classifier systems

Signal processing problems including the speaker identification problem require processing of real-valued feature vectors. Traditional cepstral encoding combined with clustering algorithms handle the closed-set speaker identification problem quite well but when it comes to the open-set problem, clustering methods show lack of performance. Furthermore, many clustering algorithms lack adaptability and the ability to learn on-the-fly. Genetic classifier systems are adaptive and they have the ability for open-ended learning. We introduce a genetic classifier system approach to the speaker identification problem and several classifier knowledge representation methods for open-set speaker identification. Experimental results show that the new system works quite well for the open-set speaker identification problem.

Text-independent open-set speaker identification for military missions using genetic rule-based system

We present a genetic classifier system approach to the text-independent open-set speaker identification problem. Classifier systems are widely used in symbolic problem for dynamically changing open-ended learning. Signal processing problems require processing of real-valued parameters that classifier systems are not designed for. On the other hand, the approaches based on common cepstral encoding with clustering algorithms handle the closed-set speaker identification quite well. This research solves the open-set problem by hybridizing these two approaches.

Front Matter: Volume 10206

This PDF file contains the front matter associated with SPIE Proceedings Volume 10206, including the Title Page, Copyright information, Table of Contents, and Conference Committee listing.

Disruptive innovation in cyber security (Conference Presentation)

The abstract is not available