natural language processing

Chaos theory finds its voice

Nathan Kahl — Tue, 19 May 2026 15:42:55 +0000

Chaos theory finds its voice Nathan Kahl Tue, 05/19/2026 - 11:42

Talk may be cheap, but it is still rich with data, and George Mason's Anomadarshi Barua is using artificial intelligence (AI) and a new approach to chaos theory to rebuild missing pieces of speech with incredible accuracy.

"When we talk, our voices have a frequency component from zero kilohertz (kHz) to around 8 kHz," said Barua, an assistant professor in the Department of Cyber Security Engineering. "Capturing all those frequencies normally requires devices to collect and process large amounts of data, which consumes power, storage, and computing resources."

He will present a paper he coauthored on this topic at the Association for Computational Linguistics (ACL) annual meeting later this summer. ACL is one of the world's leading conferences in natural language processing and speech technologies, and the paper was accepted in the top 15 percent of accepted submissions at a conference with a 19 percent acceptance rate overall.

Smaller devices and sensors operating with limited battery life or bandwidth have difficulty capturing an entire signal, and so record a smaller slice, attempting to reconstruct the missing frequencies later. Barua says this is known as bandwidth extension or bandwidth reconstruction.

Imagine listening to a song through a wall and trying to mentally fill in the muffled higher notes. Barua's research teaches AI systems to do something similar, but with far greater mathematical precision. The team's major breakthrough came from incorporating a concept not often associated with speech: chaos theory.

"Our speech is actually a chaotic signal," Barua said. "'Chaos' does not mean completely random."

Instead, he describes speech as "deterministically random," meaning sounds and phonemes are strongly connected to one another in predictable ways. "You can actually determine what could be the next phoneme if you have the sufficient information from the previous phoneme," he said.

To capture those hidden relationships, the researchers developed what Barua calls a "chaotic discriminator" inside a machine-learning framework known as a generative adversarial network, or GAN. In simple terms, one part of the AI system generates reconstructed speech while another checks whether the recreated speech preserves the natural chaotic patterns found in human voices.

The approach significantly improved reconstruction quality compared to previous methods. "We are getting more improved results compared to the previous baseline," Barua said, "because of the incorporation of the chaotic properties."

By integrating chaos-informed modeling, the team dramatically reduced the size of the AI system needed for reconstruction. "We actually reduce the size of the discriminator by 14 times," Barua said. Smaller models require less memory and computing power, making them more practical for real-world devices.

For Barua, whose earlier work focused primarily on cybersecurity before expanding into speech and natural language processing, the ACL conference's acceptance marks an important milestone. "This is my first paper in natural language processing," he said.

While the current paper focuses on speech, Barua sees much broader possibilities ahead. The same reconstruction techniques could eventually be applied to electrical signals, sonar, lidar, and other sensing technologies. "This concept is not only limited to speech," he said. "We are trying to open up a larger branch in different signal modality."

And that message comes through loud and clear.

In This Story

Anomadarshi Barua

Explore Cybersecurity at George Mason

Powering Artificial Intelligence

Topics

Cybersecurity

Department of Cyber Security Engineering

natural language processing

large language models

Artificial Intelligence

Research

College of Engineering and Computing

Using AI to uncover human smuggling networks

tdonnel — Wed, 15 Jan 2025 22:39:59 +0000

Using AI to uncover human smuggling networks tdonnel Wed, 01/15/2025 - 17:39

As vulnerable members of society, migrants are subject to exploitation. Professor Carlotta Domeniconi in the Department of Computer Science at 91�� is using artificial intelligence (AI) to combat that abuse. She is the principal investigator (PI) on a groundbreaking project aimed at understanding and modeling human smuggling networks.

Carlotta Domeniconi

In collaboration with the Department of Homeland Security’s Criminal Investigations and Network Analysis (CINA) Center for Excellence, Domeniconi and co-PIs Guadalupe Correa-Cabrera and Sean Luke are developing advanced machine learning techniques to analyze publicly available data and uncover the intricate workings of human smuggling networks between Mexico and the United States. While Domeniconi and Luke, both computer science faculty, bring expertise in machine learning, Correa-Cabrera is a professor in George Mason's Schar School of Policy and Government specializing in border studies, U.S.-Mexico relations, international security, migration studies, and illicit networks.

A significant challenge in this project is the sparsity and complexity of the data, said Domeniconi. Criminal networks are large, dynamic, and constantly evolving.

"The very nature of criminal networks—their secret nature—makes it practically impossible to gather useful data," she explained. This challenge led to the team’s innovative approach of using publicly available text data—past cases of prosecuted smugglers at both the state and federal levels in the United States as well as social media posts and news articles by and about smugglers. They are collecting data from the 1980s to 2024 to analyze the temporal dynamics of these networks.

“There are videos of smugglers actually showcasing people traveling across very difficult areas from Central America,” Domeniconi noted as an example. “You can extract a lot of information about the routes they take, where they cross the borders, and the means they use."

The team is using natural language processing (NLP) and deep learning models that are pre-trained on vast amounts of text data, enabling them to understand and extract meaningful information from the data they are collecting. Two doctoral students will support the co-PIs, one from computer science and the other from the Schar School. The team is working on automating the process of building and analyzing these criminal networks using large language models and graph mining techniques. By visualizing these networks as knowledge graphs, the team aims to mine behavioral patterns and understand how these networks adapt to changes in policy enforcement. Ultimately, this project could provide a methodology applicable to various criminal networks beyond human smuggling.

Guadalupe Correa-Cabrera

“The current project bridges the gap between machine learning and social science,” explained Correa-Cabrera. “We can replicate this same model and methodology to analyze legal cases of drug smuggling, money laundering, and arms trafficking, thereby modeling and explaining the evolution of these activities over the past few decades.”

Domeniconi's pioneering work in modeling human smuggling networks is a testament to the power of AI in addressing complex social issues. By merging domain expertise with state-of-the-art machine learning techniques, this project not only promises to unravel the clandestine operations of human smuggling but also to create a robust methodology applicable to various criminal networks. The insights gained from this research could significantly enhance national security measures, offering a strategic advantage in the fight against human smuggling and those who perpetrate this crime.

Start your George Mason journey

In This Story

Carlotta Domeniconi

Sean Luke

Guadalupe Correa-Cabrera

Interprofessional George Mason researchers awarded more than $1 million to improve outcomes for patients with depression

Jennifer Pocock — Tue, 10 Dec 2024 18:25:35 +0000

Interprofessional George Mason researchers awarded more than $1 million to improve outcomes for patients with depression Jennifer Pocock Tue, 12/10/2024 - 13:25

Researchers Farrokh Alemi and Kevin Lybarger receive 91��’s first Patient-Centered Outcomes Research Institute (PCORI) award to develop innovative Artificial Intelligence (AI) technology, including large language models, for improving antidepressant recommendations.

Farrokh Alemi and Kevin Lybarger

AI will soon receive a dose of empathy with the goal of helping to match people with depression to their best-fit medication. A team led by Farrokh Alemi, a professor in the College of Public Health (CPH), and Kevin Lybarger, an assistant professor in the College of Engineering and Computing (CEC), received $1,049,998 in research funding from the Patient-Centered Outcomes Research Institute (PCORI) to continue their work on developing an AI system that helps patients find the right depression medications.

With this funding support, Co-PIs Alemi and Lybarger will hone large language models (LLMs) to address known challenges in AI, including mitigating biases, reducing the potential for inaccurate information, and incorporating an empathetic tone, according to Alemi.

The new study will introduce an innovative way for AI to help patients make medication decisions. The AI system will engage patients in natural-language conversations to collect information about their medical history. The system will draw upon more than 10 million patient experiences with 15 different oral antidepressants and a National Institutes of Health All of Us database, which includes records from more than 80,000 participants with major depressive disorders, to help create a plan that is statistically likely to succeed. Alemi and Lybarger believe this will help alleviate the trial and error that can lead to negative patient outcomes.

The researchers will also introduce a first-of-its-kind patient simulator capable of mimicking various medical, linguistic, and behavioral characteristics. This simulator will be used to test and refine the AI system by simulating diverse patient scenarios, including infrequent but critical events such as suicidal ideation, to ensure the system’s recommendations are safe, culturally sensitive, and empathetic.

“This study was selected for its potential to address a high-priority methodological gap in patient-centered comparative clinical effectiveness research,” said PCORI Executive Director Nakela L. Cook. “We look forward to following the study’s progress and working with George Mason to share the results.”

This is the first PCORI-funded study that George Mason has received. “Depression is a major public health problem and we are excited to see the development of new AI-based decision tools, leveraging the multidisciplinary talents of our college to help tackle it,” said CPH Associate Dean of Research Alison Cuellar.

"This innovative study promises to generate methodologies for using AI for medical decision-support and for empowering patients to make critical health decisions beyond mental health,” says Özlem Uzuner, chair of CEC’s Department of Information Sciences and Technology.

This study is one of the latest funded by PCORI to examine which medical treatments work best, where and when treatment falls flat, and how to address the gaps. These studies deliver results that guide researchers in planning future studies and provide patients, their caregivers, and clinicians with the evidence-based information needed to make better-informed health and health care decisions.

In This Story

Farrokh Alemi, PhD

Kevin Lybarger

Alison Evans Cuellar, PhD, MBA

Özlem Uzuner

Research at George Mason

Topics

grants

Depression

antidepressants

natural language processing

Artificial Intelligence

Strategic Direction

CEC faculty research

Mason Momentum

Commonwealth Cyber Initiative (CCI) researchers address multidisciplinary challenges

Martha Bushong — Wed, 26 Jan 2022 20:19:34 +0000

Commonwealth Cyber Initiative (CCI) researchers address multidisciplinary challenges Martha Bushong Wed, 01/26/2022 - 15:19

In This Story

Hemant Purohit

Daniel Barbará

Vivian Genaro Motti

Whether you are an experienced software developer, a teen texting on a smartphone, or an older adult checking a bank statement, cybersecurity is part of your life. Humans and computers interact every minute of every day and cybersecurity is there to keep information safe and actions private. But normal human behavior can compromise safety and privacy.

For the next 12 months, researchers funded by the Commonwealth Cyber Initiative’s (CCI) Northern Virginia Node (NoVa Node) will be exploring the impact of human behavior on cybersecurity systems. Divided into six teams, the researchers will seek to leverage the power of their academic expertise in the social sciences, and related fields. The teams include faculty from the Colleges of Engineering and Computing, Humanities and Social Sciences, Education and Human Development, and the School of Business. Each team will explore a different aspect of the problem as they aim to translate those understandings into solutions or areas for additional investigation that can impact the welfare of Virginians.

“Human-Centric Training for Privacy and Security Controls: Bridging the Awareness Gap for Diverse Populations”

PI: Vivian Genero Motti, College of Engineering and Computing (CEC), 91��; Co-PIs: Samy El-Tawab, and Ahmad Salman, College of Integrated Sciences, James Madison University

If you retired from the workforce 25 years ago, before Wi-Fi, online shopping, banking, or smartphones, you are likely more vulnerable to cyberattacks. In fact, older adults face a disproportionate risk of suffering cyberattacks; still, they do not have access to resources and educational materials suitable to meet their needs related to human behavior and privacy protection.

Vivian Motti and her team want to do something about that. They plan to reach out to underrepresented users and characterize their level of awareness about cybersecurity. Motti and her team believe that gaining a better understanding of these populations will help inform educational content development, providing content, language, and design aspects that are all suitable to their specific user profiles.

“By adopting a user-centric design approach, this project will ensure that cybersecurity training meets users' needs for minority groups. By involving older adults front and center in the research agenda, we will establish training contents that are appropriate to their level of understanding,” says Motti. Also, besides following the training contents and retaining what they learn, they will be able to act and prevent potential attacks that could pose privacy risks.

“Impact of Human Behavior in a Mixed Traffic Environment”

PI: Linghan Zhang, CEC; Co-PIs: Nirup Menon, School of Business, Nupoor Ranade, College of Humanities and Social Sciences (CHSS),

As autonomous vehicles become more prevalent and mingle with human-driven vehicles this mixed traffic environment may comprise both. In mixed traffic, the behaviors of human drivers are unpredictable and can lead to situations that confuse autonomous vehicles and cause adverse events for both.

The CCI NoVa Node’s research in autonomous vehicles (AVs) has already garnered attention from vehicle manufacturers such as Ford, Cadillac, and Daimler-Benz. Linghan Zhang and her team aim to extend that research by studying their use in mixed traffic.

According to Linghan, the team’s goal is to reflect driving reality through a multi-vehicle simulation in mixed traffic, using driving conditions that have led to real-world collisions in the past. She says, “Prior research only focuses on a single user’s behavior, and the data collected is mainly limited to surveys and interviews. With objective driving data missing, prior experiments did not reflect on-road driving reality.”

This project could achieve valuable and meaningful data on how human driver behaviors affect other components in mixed driving environments, especially in security- and safety-critical contexts when human errors are inevitable as well as uncover what humans need to know while driving alongside AVs. The team expects that the results will be significant for autonomous vehicle implementation and policymaking.

“Towards Building Cyber-Security Resilience in a COVID-Induced Virtual Workplace”

PI: Amitava Dutta; Co-PI: Pallab Sanyal, School of Business, 91��

Before COVID-19 rocked our world, individuals and businesses were already increasing their online presence. The pandemic accelerated the speed forcing a change. People who were not comfortable in the online environment were made to go online and people who were already comfortable expanded their online presence to areas that they had previously conducted in person.

“In short, COVID-19 has caused a shift from organizational ecosystems to a virtual workplace for employees, which has opened multiple vectors for cyberattacks,” says Amitava Duta, professor at the School of Business. “Our research focuses on the behavioral and organizational aspects of cybersecurity and is motivated by the ongoing transformations following the onset of the COVID-19 pandemic.”

In their project, the team will investigate the significant changes in online behavior following the onset of the COVID-19 pandemic are. They expect their insights will help organizations build greater cyber-security resilience in a virtual workplace.

Because Washington, D.C. and Northern Virginia are home to prominent financial services organizations these businesses would have a strong interest in strengthening their cybersecurity posture to address its behavioral aspects. Soon, Amazon will also have a significant presence and retail online sales is another area frequently targeted by cybercriminals. If organizations would be willing to provide data on customer behavior on their website, the models developed from the team’s work could be refined and tailored for an important application domain.

“Characterizing and Countering User Security Fatigue in Password Enhancement through Deep Learning”

PI: Gerald Matthews, CHSS, 91��; Co-PIs: Giuseppe Anteniese and Daniel Barbará, CEC, 91��

If you already have a demanding job, you might think maintaining security is an additional burden, and not keep up with cybersecurity best practices such as updating or changing your passwords.

Professor Giuseppe Ateniese has designed a tool for enhancing password strength, based on a deep learning approach, but psychological factors may limit the adoption and impact of the tool. Everyone can be vulnerable to security fatigue and lax cybersecurity practices can have major societal consequences—threats to national security, financial losses to individuals and organizations, and invasion of privacy.

Introducing security tools powered by Artificial Intelligence, when successful, will counteract typical human fallibilities and promote safety in computer systems across government, industry, and personal use. This project investigates the effect of security fatigue on the use of Anteniese’s tool. It will also explore strategies for mitigating fatigue and supporting user engagement.

The team believes that enhancing employees' ability and motivation to maintain effective security protocols has immediate economic benefits and the research has the potential to suggest design features of security tools that can support commercialization as well as training protocols.

“Enabling Invisible Security and Privacy for Resilient Human-Centric Cybersecurity Systems”

PI: Eric Osterweil, CEC, 91��; Co-PI: Matt Canham, CHSS, 91��

For decades, cryptography has been one of cybersecurity’s most essential tools. While its utility is certain, its complexity limits its use for non-experts. The result—non-experts fall prey to cybercriminals for many reasons including lack of knowledge, incorrect thought processes, and the inability to invest adequate time and resources to implement proper data protection.

Eric Osterweil and his co-investigator Matt Canham hope to change that through their work with the CCI NoVa Node. “This project will seed a critical foundation for adaptive cybersecurity protections for human users’ end-to-end encryption (E2EE) needs. The results from this project will be used as foundations for enhancing a core staple of Internet communications (email) and future advances in prescriptive protections for Cybersecurity Threat Intelligence (CTI) information sharing,” says Osterweil.

The CTI industry continues to grow, with companies, federal agencies, and international communities relying on CTI. In Virginia, where federal agencies and their partners routinely conduct transactions over email, this is especially true. Their view is that building human usable E2EE protections and extending those to adaptive CTI will be directly relevant to operational cybersecurity projects and needs throughout the industry and public sectors in Virginia.

The pair believes that a key benefit to the Commonwealth will include course-related exposure of this material to the students at 91��. “Students will be able to showcase both the results of this work and their own derived qualifications to benefit their entry into local industry and jumpstart their ascension to professional careers,” says Osterweil.

"Characterizing Biases in Automated Scam Detection Tools for Social Media to Aid Individuals with Developmental Disabilities"

PI: Hemant Purohit, CEC; Co-PIs: Géraldine Walther, CHHS; Matt Peterson, CHHS; YooSun Chung, CEHD

Designers of scam detection tools often focus on improving the computational accuracy of the methods, especially those with state-of-the-art Natural Language Processing (NLP) and Machine Learning (ML)-based techniques, but their understanding of the diverse human behavior can be limited. This project aims to build a foundation for inclusive cybersecurity technologies to protect individuals with disabilities from online scams using a unique interdisciplinary collaborative approach between computing and non-computing researchers.

Specifically, the team’s objective is to uncover the biases in the existing scam detection techniques for social media using NLP and ML methods. “We will conduct Eye Tracking analyses using a labeled scam dataset of social media posts from existing literature on online cybersecurity and study the differences between the attention patterns of individuals with and without developmental disabilities when perceiving scam posts,” says Hemant Purohit.

The project hopes to gain insights that will support cybersecurity training development for reducing online fraud for individuals with special education needs. At the same time, the researchers want to identify limitations in automated scam detection tools and help create more effective cybersecurity tools that can protect user groups in our communities.

Topics

Cybersecurity

Commonwealth Cyber Initiative (CCI)

cryptography

coronavirus; covid-19

Social Media

natural language processing

Research

Artificial Intelligence

Computer science expert using natural language processing to improve equality in language technologies

Anonymous — Wed, 17 Feb 2021 14:24:56 +0000

Computer science expert using natural language processing to improve equality in language technologies Anonymous (not verified) Wed, 02/17/2021 - 09:24

Computer science researcher Antonis Anastasopoulos uses his love for computer science, language, and linguistics to improve equality in language technologies.

When people ask Siri, Alexa, or Google Assistant a question, they expect the programs to understand them, but that is not always the case, he says.

Antonis Anastaspopoulos, photo provided.

A person’s language, accent, dialect, and even gender can have an impact, preventing the system from interpreting them correctly, says Anastasopoulos, an assistant professor in the Department of Computer Science and an expert in natural language processing, which is how computers attempt to process and understand human languages.

“The systems don’t work equally well for everyone,” says Anastasopoulos, who speaks Greek (his native language), English, German, Swedish, Italian, and some Spanish.

He is one of several co-principal investigators who received a new National Science Foundation-Amazon grant for their research, “Quantifying and Mitigating Disparities in Language Technologies.”

In the fall, Anastasopoulos also won a Google 2020 Award for Inclusion Research for a project on how accent and dialect impact language technologies.

For the NSF grant, he and experts from Carnegie Mellon University and the University of Washington are studying areas where there is bias in language technologies and measuring the discrepancies. Then they will attempt to mitigate the inequalities.

“We want to measure the extent to which the diversity of language affects the utility that speakers get from language technologies,” Anastasopoulos says. “We will focus on automatic translation and speech recognition since they are perhaps the most commonly used language technologies throughout the world.”

His research will apply to all languages. It’s important to look deeply into languages for differences because languages are flexible and diverse, he says. “There are many regional variations that are different from the standard.”

He also recently received a $350,000 grant from the National Endowment for the Humanities (NEH) to build optical character recognition tools to convert scanned images of text to a machine-readable format for endangered languages.

“We are working on training machine-learning models to process images and texts in the books and documents of indigenous languages from central and South America so that these works can be made accessible to everyone,” he says. “We are building technologies to study those languages computationally.”

Anastasopoulos is also part of a prestigious group of machine-translation researchers, including experts from Facebook, Google, Amazon, and Microsoft, who are creating automatic tools that translate COVID-19-related content for communities where people don’t speak the languages most often used by large health organizations, including the World Health Organization.

“We are working closely with Translators without Borders. So far, we have produced terminologies for more than 200 languages and a large dataset for 35 languages, some of them extremely under-served by the current solution.”

Topics

computer science

computing

natural language processing

natural language processing

Chaos theory finds its voice

In This Story

Related Stories

Topics

Using AI to uncover human smuggling networks

In This Story

Related News

Interprofessional George Mason researchers awarded more than $1 million to improve outcomes for patients with depression

In This Story

Related Stories

Topics

Commonwealth Cyber Initiative (CCI) researchers address multidisciplinary challenges

In This Story

Topics

Computer science expert using natural language processing to improve equality in language technologies

Topics