Research
Towards Equitable Privacy-Preserving Machine Learning Algorithms.
(still in progress) Collaborators: Dr. Ross Maciejewski (ASU), Patrick Valente (UNCP), Aurelio Medina (UNCP), Tiankai Xie (ASU), Jiayi Hong(ASU), Yixuan Wang(ASU)
This study explores the research related to the interconnections between privacy and fairness in Machine Learning (ML) algorithms. As ML technologies become more pervasive in society, the need for reliable, secure, and unbiased ML models becomes critical. Unfortunately, numerous real-world incidents have been reported involving privacy breaches and biased decisions made by ML systems. ML algorithms make decisions based on training data, and any biases or imbalances present in the data can lead to unfair outcomes. Moreover, when developing ML models for people-oriented tasks, safeguarding the privacy of individuals who contribute their data for training and testing the model is crucial. The effects of privacy and fairness enhancements on model accuracy have been extensively studied, but the majority of these studies have treated privacy and fairness as separate concerns. A limited number of studies explore their interconnections and the potential for simultaneous enhancement. In this survey, we examined eighty-seven papers, out of which twenty-six addressed privacy and fairness in combination. To provide a comprehensive overview, we included a table summarizing the types of privacy and fairness concepts addressed in each paper and their key findings. Notably, most of the surveyed papers focused on exploring how changes in privacy impact fairness, not the reverse relationship. This suggests a crucial research gap that requires attention. In conclusion, this paper outlines potential future research directions, aiming to foster a deeper understanding of the intricate interplay between privacy and fairness in ML algorithms.
equitable privacy preserving ML system-survey.pdfTowards Equitable Privacy-Preserving Machine Learning Algorithms 07-26-23.pptxTowards Creating Learning Objects for Cyber Security Education
Colloborators: Shan Suthaharan (UNCG), Gregory Ross (SPAWAR, Charlseton)
ABSTRACT: Platform-independent encapsulated and self-contained labs can enhance the delivery of cybersecurity teaching modules in resource-limited academic institutions. Cybersecurity teaching modules usually incorporate hands-on labs using tools such as NETLAB+, NS-3, SEED, GENI, etc. There are two major problems in using such labs; students need to learn the underlying tool before doing labs, and setting-up a lab may require significant IT support and resources. In addition, development and delivery of security labs is barricaded by two other technical realities: built-in security controls in computer systems and networks (e.g. NX-bits, ASL-Randomization, Canaries, Firewalls, etc.), and heterogeneity in students-owned operating environments. In our proposed work, we have studied the concept of container-based educational solutions by leveraging the unique features of Docker containers (i.e., light-weight packaging of code and all of its system dependencies) to build platform-independent encapsulated cybersecurity labs. The benefits of our approach includes, cost-effective teaching environment, scalability over heterogeneous operating platforms, independence from hardware, software, and network controls, and hassle-free learning experience for students. As a proof of concept, we have developed an interoperable encapsulated lab for giving hands-on experience to students in one aspect of cybersecurity, buffer overflow. The lab was created using Docker and uploaded to a central repository. Students can pull this lab from anywhere, and can execute it on any platform that runs Docker daemon. We will use the knowledge and experience gained in this study to build our future research in the areas of building Learning Objects for cybersecurity education.
Opening Gaming Logic for Teaching First Programming Language
ABSTRACT: Learning first computer programming language is considered very challenging. Research shows that intrinsic motivation plays key role in learning complex disciplines such as programming. In this study, we are examining a novel strategy to increase the intrinsic motivation of the novice programmers. Digital games are created using programming languages. Young students are fond of playing digital games, and they are also interested in creating their own games. This research tries to capitalize both of these availing passions as motivating factors to learn programming. We are creating educational games for learning programming, and in particular, we intentionally include a number of bugs in the game program. Students will be challenged to fix the bugs while they play. In this strategy, students are not only entertained but also challenged to debug the games, and therefore, their self-esteem will be nurtured. We believe, this Challenge Based Learning strategy will intrinsically motivate the students to actively engage in learning while playing. A prototype of a simple casual game was created to conduct a pilot study. The preliminary results were encouraging. Currently, two small educational games are being created in order to teach conditional branching and iterative looping concepts. These games will be used in real classroom settings to evaluate the effectiveness of the proposed strategy. This poster will describe the importance of this research, relevant past research, proposed approach in this study, key features in the game design, and evaluation methodologies. The prototype will also be demonstrated.
Enhancing Intrusion Detection Algorithms in Software Defined Networks: Towards Fuzzy Logic & Bayesian Network based Solution
Colloborator: Gregory Ross (SPAWAR, Charleston)
ABSTRACT: Traditionally, Intrusion Detection Systems (IDS) are being used to detect suspicious packets in the computer networks and raise alarms to the network administrators. However, the contemporary IDSs have a main drawback, their false-positive rates are intolerably high. Incidently, Software Defined Networks (SDN) are becoming more and more popular due to rapidly changing network usage requirements. In this research, we are investigating how IDSs in SDN environments can be augmented by Fuzzy Logic and Bayesian Network techniques in order to reduce their false-positive alarm rates. We are proposing an approach where a dynamic network flow model will be maintained at the controller. The flow model will be continuously updated using Bayesian Network Theory.. In addition, a traditional IDS such as Snort will be used to provide an initial risk estimation of a forwarded packet. Based on the current flow model and the risk estimation, the proposed fuzzy algorithm will suggest a suitable action, whether to drop the packet, forward it, or send it to further investigation (alarm). We believe that our approach will significantly reduce the false-positive alarms in Intrusion Detection Process in SDN environments. The proposed approach will be evaluated using Mininet (a network emulation environment). Various network attacks will be simulated using Python programs. The risk of a packet will be estimated by a traditional IDS running on an SDN controller (e.g.: RYU + Snort). The proposed Bayesian and Fuzzy algorithms will be implemented. The statistical significance of changes in false-positive (and false-negative) rates will be analyzed and reported.
A Multi-factor Authentication Technology for Self-Service Password Resetting using Mobile Camera Fingerprints, Bio-metrics and Spatio-Temporal Information
ABSTARCT: The objective of this research is to identify secure, user-friendly, and cost-effective technologies for Self-Service Password Reset systems. Despite the universally acknowledged fact that username-password model of authentication is the key cause for many notorious cyber space attacks, researchers predict that password will survive as a prime authentication tool for a foreseeable future. Memorizing complex passwords for a number of accounts will be a cognitive burden for the average person. Therefore, people tend to choose familiar and easily memorable texts as their passwords. Since these passwords can be easily compromised, on-line systems encourage (or enforce) the users to choose lengthy random passwords. To cope with this pressure, many users select complicated passwords and then forget them frequently. As a remedy, web applications provide various password reset options. Research shows that many popular Self-Service Password Reset systems are vulnerable to various security breaches. This study will analyze the possibilities of designing a multi-factor authentication model for self-service password resetting process. The proposed technology will utilize camera finger prints of the mobile-phones along with biometric and spatio-temporal information. It has been found that every mobile camera has a unique digital sensor pattern noise known as camera finger print. A prototype of a self-service password reset system that incorporates the proposed multi-factor authentication model is being created in order to evaluate the effectiveness of the proposed approach.
ABSTARCT: Multiple choice (MC) tests are typically used for formative assessments in eLearning. However, the interaction bandwidth of traditional MC tests is severely limited; students can only make a tick on an answer option. Therefore the eLearning system needs to make assumptions to fill-in the gaps. Nevertheless, the system’s assumption may be wrong. A student may guess a correct answer or may make a careless mistake. If an eLearning system often makes stereo-typed pedagogical decisions based on wrong assumptions, eventually, the students would get frustrated. This research describes an artificial intelligent based methodology that could provide individualized test sequencing and personalized feedback.