UMBC Computer Science and Electrical Engineering Department

Permanent URI for this collectionhttp://hdl.handle.net/11603/50

The Computer Science and Electrical Engineering Department aims to maintain a program of excellence in teaching, research, and service for all of its programs. At the undergraduate level, we will provide students with a firm foundation of both the theory and practice of Computer Science and Computer Engineering. Our curricula also give students the social, ethical, and liberal education needed to make significant contributions to society. Students receiving a bachelor’s degree are ready to enter the work force as productive computer scientists or computer engineers, or to continue their education at the graduate or professional level.

At the graduate level, we are committed to developing the research and professional capabilities of students in Computer Science, Computer Engineering, Electrical Engineering and Cybersecurity. Our programs provide a deeper mastery of the basics of these fields, as well as opportunities to collaborate on leading-edge research with our faculty. Our faculty are engaged in both practical and theoretical research, often in partnership with government agencies, private industry and non-governmental organizations. The aim of this research is to advance knowledge within our disciplines and also to contribute to solving problems faced by our society.

Browse

Now showing 1 - 20 of 2189

ADAPTIVE IE: Investigating the Complementarity of Human-AI Collaboration to Adaptively Extract Information on-the-fly
(ACL, 2025-01) Mondal, Ishani; Yuan, Michelle; N, Anandhavelu; Garimella, Aparna; Ferraro, Francis; Blair-Stanek, Andrew; Van Durme, Benjamin; Boyd-Graber, Jordan
Information extraction (IE) needs vary over time, where a flexible information extraction (IE) system can be useful. Despite this, existing IE systems are either fully supervised, requiring expensive human annotations, or fully unsupervised, extracting information that often do not cater to user`s needs. To address these issues, we formally introduce the task of “IE on-the-fly”, and address the problem using our proposed Adaptive IE framework that uses human-in-the-loop refinement to adapt to changing user questions. Through human experiments on three diverse datasets, we demonstrate that Adaptive IE is a domain-agnostic, responsive, efficient framework for helping users access useful information while quickly reorganizing information in response to evolving information needs.
Single Image Super Resolution Using AI Generated Images
(2025-01-18) Singh, Amanjot; Khan, Faisal Rasheed; Singh, Mrinalini
Image super-resolution has become increasingly important in various applications because of their demand for producing high output images from the low input images. Earlier for the image enhancements techniques like deblurring were performed to get the quality image. With the advancements in the Generative Adversarial Networks (GAN), the generating of high-quality image from the low-quality image has been outstanding. The models like SRGAN, ESRGAN [12]are the competitive models which make the Image-Resolution look good because of their performance on the images. But the architecture of the SRGAN which is a state-of-art model is complex and ESRGAN is built on the SRGAN, but by observing the results of the SRGAN the image quality looks good. We try to build a Super-Image Resolution by having the less complex architecture which is faster than SRGAN and the results aren’t compromising even after reducing the architecture complexity. We have built our base model based on the SRGAN by reducing the complexity in the architecture. In our final model we added another discriminator layer which enhances the sub parts of the images to improve the image quality. Our aim is to build an efficient model where the architecture of our model is less complex than SRGAN [14]and give as competitive results as SRGAN. Our results for the final model compared to our base model shows that there were significant improvements in the image quality. The code link for our project is here:https://github.com/faisalkhansk3283/ Computer_Vision_Extended_SRGAN
Development and Initial Testing of XR-Based Fence Diagrams for Polar Science
(IEEE, 2023-07) Tack, Naomi; Holschuh, Nicholas; Sharma, Sharad; Williams, Rebecca M.; Engel, Don
Earth’s ice sheets are the largest contributor to sea level rise. For this reason, understanding the flow and topology of ice sheets is crucial for the development of accurate models and predictions. In order to aid in the generation of such models, ice penetrating radar is used to collect images of the ice sheet through both airborne and ground-based platforms. Glaciologists then take these images and visualize them in 3D fence diagrams on a flat 2D screen. We aim to consider the benefits that an XR visualization of these diagrams may provide to enable better data comprehension, annotation, and collaborative work. In this paper, we discuss our initial development and evaluation of such an XR system.
Visualizing the Greenland Ice Sheet in VR using Immersive Fence Diagrams
(ACM, 2023-09-10) Tack, Naomi; Williams, Rebecca M.; Holschuh, Nicholas; Sharma, Sharad; Engel, Don
The melting of the ice sheets covering Greenland and Antarctica are primary drivers of sea level rise. Predicting the rate of ice loss depends on modeling the ice dynamics. Ice penetrating radar provides the ability to capture images through the ice sheet, down to the bedrock. Historical environmental and climate perturbations cause small changes to the dielectric constant of ice, which are visually manifested as layers of varying brightness in the radar imagery. To understand how the flow of ice has progressed between neighboring image slices, glaciologists use Fence Diagrams to visualize several cross-sections at once. Here, we describe the immersive virtual reality (VR) fence diagrams we have developed. The goal of our system is to enable glaciologists to make sense of these data and thereby predict future ice loss.
Signal Processing of Images for Convective Boundary Layer Height Estimation from Radar (SPICER) and multi-instrument verification
(IEEE, 2025-01-13) Porta, Delia Tatiana Della; Demoz, Belay
The study of the planetary boundary layer (PBL) is one of the main topics of the atmospheric community. The current study presents a new algorithm for PBL height determination using a publicly available but unexplored data source, the Weather Service Radar (WSR-88D). The diurnal evolution of the PBL is also known as Convective Boundary Layer (CBL), key in the study of convection and precipitation. This paper presents the Signal Processing of Images for Convective Boundary Layer Height Estimation (SPICER) algorithm that can automatically detect the CBL Height (CBLH) for all of the 159 radar locations across the United States during clear days. The present work is the first step to applying SPICER to a network of Next Generation Radars (NEXRAD) with continuous countrywide coverage. With the possible combination with the Automated Surface Observing System network (ASOS), a source of ceilometer profile data, a validated dataset of CBLH estimates can be expected soon. The algorithm treats averaged differential reflectivity vs range as an image and applies filtering plus Canny edge detection to estimate the CBLH. In addition, another algorithm is presented to automate the detection of the mixing layer height (MLH), a proxy for CBLH from Raman Lidar and a 915 MHz wind profiler. A comparison of CBLH estimates vs widely used methods in meteorology (Radiosondes, Raman Lidar, ceilometer, 915 MHz wind profiler, and Doppler Lidar-based derived Value-Added Product (VAP) ) is performed to validate the NEXRAD detected CBLH using SPICER. The SPICER algorithm shows over 0.9 correlation with radiosonde measurements.
Mapping the Edges of Mass Spectral Prediction: Evaluation of Machine Learning EIMS Prediction for Xeno Amino Acids
(2025-01-14) Brown, Sean M.; Allgair, Evan; Kryštůfek, Robin
Mass spectrometry is one of the most effective analytical methods for unknown compound identification. By comparing observed m/z spectra with a database of experimentally determined spectra, this process identifies compound(s) in any given sample. Unknown sample identification is thus limited to whatever has been experimentally determined. To address the reliance on experimentally determined signatures, multiple state-of-the-art MS spectra prediction algorithms have been developed within the past half decade. Here we evaluate the accuracy of the NEIMS spectral prediction algorithm. We focus our analyses on monosubstituted α-amino acids given their significance as important targets for astrobiology, synthetic biology, and diverse biomedical applications. Our general intent is to inform those using generated spectra for detection of unknown biomolecules. We find predicted spectra are inaccurate for amino acids beyond the algorithms training data. Interestingly, these inaccuracies are not explained by physicochemical differences or the derivatization state of the amino acids measured. We thus highlight the need to improve both current machine learning based approaches and further optimization of ab initio spectral prediction algorithms so as to expand databases for structures beyond what is currently experimentally possible, even including theoretical molecules.
MedGrad E-CLIP: Enhancing Trust and Transparency in AI-Driven Skin Lesion Diagnosis
(IEEE, 2025-01-12) Kamal, Sadia; Oates, Tim
As deep learning models gain attraction in medical data, ensuring transparent and trustworthy decision-making is essential. In skin cancer diagnosis, while advancements in lesion detection and classification have improved accuracy, the black-box nature of these methods poses challenges in understanding their decision processes, leading to trust issues among physicians. This study leverages the CLIP (Contrastive Language-Image Pretraining) model, trained on different skin lesion datasets, to capture meaningful relationships between visual features and diagnostic criteria terms. To further enhance transparency, we propose a method called MedGrad E-CLIP, which builds on gradient-based E-CLIP by incorporating a weighted entropy mechanism designed for complex medical imaging like skin lesions. This approach highlights critical image regions linked to specific diagnostic descriptions. The developed integrated pipeline not only classifies skin lesions by matching corresponding descriptions but also adds an essential layer of explainability developed especially for medical data. By visually explaining how different features in an image relates to diagnostic criteria, this approach demonstrates the potential of advanced vision-language models in medical image analysis, ultimately improving transparency, robustness, and trust in AI-driven diagnostic systems.
Comparison of Several Neural Network-Enhanced Sub-Grid Scale Stress Models for Meso-Scale Hurricane Boundary Layer Flow Simulation
(AIAA, 2025-01-03) Hasan, MD Badrul; Yu, Meilin; Oates, Tim
The complicated energy cascade and backscatter dynamics present a challenge when studying turbulent flows in storms at the meso-scale. When performing standard large-eddy simulations (LES), sub-grid scale (SGS) stress models usually fail to consider energy backscatter. These models assume that kinetic energy only moves continuously from larger to smaller scales. However, coherent energy backscatter structures exist when analyzing hurricane boundary layer flows at the meso-scale. Our recent research has shown that machine-learning SGS models trained with high-resolution data can effectively forecast forward and backward energy transfers in meso-scale hurricane-like vortex flows. Therein, physical and geometrical invariances were introduced to better represent flow physics. This further improved the predictability and generalizability of machine-learning-enhanced SGS models. In this study, we compare the performance of several machine-learning-enhanced SGS models, especially those based on neural networks (NNs), with varying physical and geometrical invariance embedding levels for SGS stress modeling in an a priori sense, which sets the cornerstone for ongoing a posteriori tests of NN models.
Examining Engagement with Disinformation Accounts on Instagram Using Web Archives
(2025-01-10) Prince, Leah; Weigle, Michele C.
Disinformation poses a serious threat to society. Researchers have shown that many people spread disinformation, especially in relation to the COVID-19 pandemic, on social media platforms like Instagram. However, performing disinformation analysis is difficult if those users have been banned. Previous research has shown that web archives are useful for performing disinformation analysis on pages that are no longer live, but this approach is limited by web page capture availability. To better understand how disinformation is spreading on Instagram, we examine how disinformation actors are utilizing hashtags and account mentions to boost engagement by extracting engagement metrics from archived Instagram account pages. We use the data gathered from these archived webpages, or mementos, to perform network analysis showing how engagement connects users. We then perform clustering based on hashtag frequency to examine how people searching for reputable content are being exposed to disinformation by identifying groupings that contain both health authority accounts and anti-vax accounts. Our findings indicate that roughly one-fifth of intra-group hashtags are also distributed inter-group. Limited memento availability remains an obstacle to comparing reputable accounts to disinformation accounts, but a higher percentage of Instagram account page mementos from the Archive.today web archive are able to be scraped than those from the Internet Archive’s Wayback Machine.
Water Flow Detection Device Based on Sound Data Analysis and Machine Learning to Detect Water Leakage
(2025-01-19) Pourmehrani, Hossein; Hosseini, Reshad; Moradi, Hadi
In this paper, we introduce a novel mechanism that uses machine learning techniques to detect water leaks in pipes. The proposed simple and low-cost mechanism is designed that can be easily installed on building pipes with various sizes. The system works based on gathering and amplifying water flow signals using a mechanical sound amplifier. Then sounds are recorded and converted to digital signals in order to be analyzed. After feature extraction and selection, deep neural networks are used to discriminate between with and without leak pipes. The experimental results show that this device can detect at least 100 milliliters per minute (mL/min) of water flow in a pipe so that it can be used as a core of a water leakage detection system.
SONDAR: Size and Shape Measurements Using Acoustic Imaging
(ACM, 2024-10-01) Liang, Xiaoxuan; Wei, Zhaolong; Li, Dong; Xiong, Jie; Gummeson, Jeremy
Acoustic signal has been used to sense important contextual information of targets such as location and movement. This paper explores its potential for measuring the geometric information of moving targets, i.e., shape and size. We propose SONDAR, a novel shape and size measurement system using Inverse Synthetic Aperture Radar (ISAR) imaging on commodity devices. We first design a Doppler-free echo alignment method to accurately align target reflections even if there exists a severe Doppler effect. Then we down-convert the received signals reflected from multiple scatter points on the target and construct a modulated downchirp signal to generate the image. We further develop a lightweight approach to extract the geometric information from a 2-D frequency image. We implement and evaluate a proof-of-concept system on both a Bela platform and a smartphone. Extensive experiments show that we can correctly estimate the target shape and achieve a millimeterlevel size measurement accuracy. Our system can achieve high accuracies even when the target moves along a deviated trajectory, at a relatively high speed, and under obstruction.
NANOFABRICATION AND TESTING OF PHOTONICS AND ELECTRONICS DEVICES
(2024-01-01) Sood, Rachit Mohan; Choa, Fow-Sen; Computer Science and Electrical Engineering; Engineering, Computer
Nanotechnology, with its ability to manipulate matter at the smallest scales has become the cornerstone of modern optoelectronics. One nanometer is equivalent to one billionth or 10-9 of a meter. To put the scale into perspective, the size of a nanometer compared to a meter is similar to the size of a marble compared to the earth. Nanofabrication, arises from nanotechnology, focuses on the methods to build the nanometer scale components, and features with precise accuracy. Through advancements in nanotechnology, photonic devices are tailored to achieve the remarkable control over light at nanoscale dimensions which includes nanostructures for antireflection coatings to enhance the transmission and planar meta lenses to precisely focus the light at certain focal length. Antireflection coatings are used to suppress reflection and increase the optical transmissions, however, these coatings whether it is single layer, or multi-layer cannot withstand the harsh environmental conditions and generally suffers from degradation which deteriorate its optical performances. Inspired from moth eye, subwavelength structures designed periodically were fabricated on the III-V materials, specifically gallium arsenide, to increase its transmittance in the mid-infrared (mid-IR) range wavelengths. Nanopillars structures were fabricated using the top-down approach to pattern the structures directly on the substrate. By varying the pitch of nanostructures, further improvement in the transmission was measured and the results were validated using the rigorous coupled wave analysis (RCWA) model. Similar nanostructures but much smaller in size were written using the E-beam lithography technique on polysilicon coated quartz sample to focus and reflect the light back at the incident direction. The meta lens was first designed by choosing the phase profiles and simulating the different size nano atoms to generate the transmission and phase values chart. Mapping the phase profile against the correct position of meta-atoms helps to produce the desired delays across the lens.Electronic devices, on the other hand, manipulate the movement of electrons instead of light to process the information. Traditional electronic devices such as MOSFETs and BJTs suffers from charge carrier and low electron mobility which makes them less preferred choice for high frequency applications. High electron mobility transistors or simply HEMTs, which possess reduced scattering and high electron mobility, are therefore used for advanced microwave and millimeter applications. Herein, I have also fabricated InP based HEMTs for the terahertz image detection utilizing the techniques such as MBE, PVD, wet etching and lift off.
THE IMPACT OF DEVICE AGING ON THE RESILIENCY OF CRYPTOGRAPHIC DEVICES AGAINST PHYSICAL ATTACKS
(2024-01-01) ANIK, MD TOUFIQ HASAN; Karimi, Naghmeh Dr.; Computer Science and Electrical Engineering; Engineering, Computer
Cryptographic chips offer continued advances in authenticating messages and devices as well as preserving the integrity and confidentiality of sensitive information through the implementation of cryptographic algorithms in hardware. These pieces of silicon combine the benefits of cryptographic applications with the speed and power advantages of hardware implementations. Indeed, cryptographic devices are used as an essential entity in almost all systems that deal with sensitive data, e.g., banking, military, transportation, medical, internet-of-things, and cloud networks. Disruption on these devices can be forced by adversaries to uncover secret information. In fact, the probability of such devices being attacked or hacked is rapidly growing; necessitating the protection of such devices against attacks and in turn secret data leakage. One such attack that can be launched on cryptographic devices is leaking the secret key through side-channel analysis. Indeed, while operating, cryptographic devices leave traces of side-channel information such as power consumption, electromagnetic emanation, running time, and so on. This information can be utilized to conduct statistical analysis, the so-called Side-Channel Attacks (SCA), to retrieve secret information. As mentioned, although cryptographic cores have been developed to maintain security and trust, their physical implementation can be compromised by the adversaries who aim at extracting the sensitive information these chips conceal. Thereby, it is essential to assure the security of the sensitive tasks performed by these circuits and to guarantee the security of information stored within these devices. Security challenges for cryptographic devices can range from the attacks launched to directly retrieve sensitive data (e.g., Side-Channel Analysis attacks and Fault-Injection at tacks) to the attacks in which a cryptographic design is tampered via inserting hardware Trojans to ease information leakage or to cause a denial of service. Protecting cryptographic devices against all such attacks is a major security concern.Side-channel analysis attack via analyzing the device’s power consumption is one of the most popular form of attacks threatening cryptographic devices. Power analysis attacks are carried out by investigating the data being processed as well as the device’s associated power usage. In contrast to power analysis attacks, in the fault injection attacks the adversary has a more active role; injecting faults in the targeted chip to facilitate extracting sensitive data based on a comparison or correlation between the faulty and golden outputs. A wide number of countermeasures have been proposed in recent years to thwart power analysis and fault injection attacks. However, the resiliency of such countermeasures may vary in different operating conditions such as operating voltage, temperature, as well as aging-related degradation. The degradation imposed by aging is considered inevitable for electronic circuits. Therefore, it is highly important to investigate the security of devices that have been used for a while, i.e., have been aged, to ensure their security is not compromised when the devices are aged. This research focuses on the impact of device aging on the security of crypto graphic devices against side-channel analysis attacks, in particular power analysis attacks as well as the resiliency of such devices against fault injection attacks. Accordingly, we first investigated the impact of device aging on the resiliency of unprotected cryptographic devices when subject to profiling and non-profiling power analysis attacks. Then we moved one step forward to analyze the aging impact on the existing countermeasures against power analysis attacks. These countermeasures are mainly classified into two groups: hiding and masking. In our investigation, we conducted power analysis attacks on hiding-protected devices when new and when aged. In particular, we target Sense Amplifier Based Logic (SABL) and Wave Dynamic Differential Logic (WDDL) circuits as the two main existing hiding countermeasures. We showed that, the success rate of the power analysis attacks increases in both SABL- and WDDL-protected circuits over time, i.e., their security diminishes over time. To address the problem, we propose an aging-resilient variation of the SABL circuit to ensure long-lasting security. We also investigated the impact of device aging on several state-of-the-art masking countermeasures tailored to protect against power analysis attacks. The results showed that the protection offered by the state-of-the-art masked devices fluctuates with aging. This may potentially expose the device to vulnerabilities as it ages. Finally, we validated our findings on real silicon, specifically on FPGA. In order to protect cryptographic devices against fault injection attacks, we developed a digital-sensor based failure-detection framework, devised an efficient characterization methodology for the considered sensor, and proposed an aging aware dimensioning algorithm to optimize the sensor hardware based on the whole range of operating conditions. We showed the efficiency of the proposed framework in detecting fault attacks induced by change of temperature, voltage, and clock frequency. These findings were also evaluated using platforms. In sum, our solutions provide long lasting security for cryptographic devices against physical attacks and thus highly benefits the industry and government sectors.
DUNE (Deep UNet++ Ensemble): A Reanalysis-Driven AI-based Climate Forecasting Framework for Monthly, Seasonal, and Annual Predictions
(2024-01-01) Shukla, Pratik Ketanbhai; Halem, Milton; Computer Science and Electrical Engineering; Computer Science
Capitalizing on the recent availability of ERA5 monthly averaged long-term data records of mean atmospheric and climate fields based on high-resolution reanalysis, deep-learning architectures offer an alternative to physics-based daily numerical weather predictions for subseasonal to seasonal (S2S) and annual means. A novel Deep UNet++ Ensemble (DUNE) neural architecture is introduced, employing multi-encoder-decoder structures with residual blocks. When initialized from a prior month or year, this architecture produced the first AI-based global monthly, seasonal, or annual mean forecast of 2-meter temperatures (T2m) and sea surface temperatures (SST). ERA5 monthly mean data is used as input for T2m over land, SST over oceans, and solar radiation at the top of the atmosphere for each month of 40 years to train the model. Validation forecasts are performed for an additional two years, followed by five years of forecast evaluations to account for natural annual variability. AI-trained inference forecast weights generate forecasts in seconds, enabling ensemble seasonal forecasts. Root Mean Squared Error (RMSE), Anomaly Correlation Coefficient (ACC), and Heidke Skill Score (HSS) statistics are presented globally and over specific regions. These forecasts outperform persistence, climatology, and multiple linear regression for all domains. DUNE forecasts demonstrate comparable statistical accuracy to NOAA’s operational monthly and seasonal probabilistic outlook forecasts over the US but at significantly higher resolutions. RMSE and ACC error statistics for other recent AI-based daily forecasts also show superior performance for DUNE-based forecasts. The DUNE model’s application to an ensemble data assimilation cycle shows comparable forecast accuracy with a single high-resolution model, potentially eliminating the need for retraining on extrapolated datasets.
Shigure – A high-performance object storage system with rich metadata provided by a distributed ledger
(2024-01-01) Sebald, Lawrence John; Yesha, Yaacov; Computer Science and Electrical Engineering; Computer Science
Networked computer systems often require massive data storage systems for powering the input needs of the system and providing for a place to store computation results. However, not all networked systems have such high data demands -- sometimes a networked system is just one that holds a family's media and data backups. No matter what the size of the system, a need for data storage will still exist. Where there is data, there is also metadata to describe and help search through the data. Metadata management within networked data storage is a messy proposition, and is often a woefully overlooked component of the system. Typical computer filesystems provide little in the way of metadata support beyond basic file attributes without relying on external software to maintain some sort of index for the user. Modern object storage systems do provide additional metadata support, but do not provide any sort of indexing beyond what a typical filesystem does. I have developed, implemented, and tested Shigure to address the problems of providing adaptable storage with rich metadata support, including indexing. Shigure combines a modern, high-performance object storage system with a distributed ledger for keeping track of and indexing metadata and providing access control to the data stored within the system. This allows Shigure to provide scalable storage with rich metadata indexing support that allows unique access patterns that typical filesystems and object storage systems do not provide without the need for an external database. Shigure also implements a hierarchical model for user permissions that goes beyond the typical idea of users and groups that filesystems provide, allowing for a user-friendly way of sandboxing client programs that wish to make use of the data stored within the system. Shigure has been tested on a variety designs of storage systems to evaluate its performance in various situations -- from a single-node system that might be used by an individual in a home network to a system approximating a big data cloud storage environment with multiple storage nodes communicating over the Internet.
Reg-Tune: A Regression-focused Fine-tuning Technique for Energy-efficient Embedded Deployment
(2024-01-01) Mazumder, Arnab Neelim; Younis, Dr. Mohamed; Mohsenin, Dr. Tinoosh; Computer Science and Electrical Engineering; Engineering, Computer
Fine-tuning deep neural networks (DNNs) is essential for developing inference modules suitable for deployment on edge or FPGA (Field Programmable Gate Arrays) platforms. Traditionally, various parameters across DNN layers have been explored using grid search, neural architecture search (NAS), and other brute-force techniques. While these methods yield optimal network parameters, the search process can be time-consuming and may not account for the deployment constraints of the target platform. The dissertation addresses this issue by formulating hardware-aware regression polynomials for the energy-efficient deployment of DNN models through Reg-Tune. A general formulation is provided to profile different metrics across various device platforms for both single and multi-objective solutions. The objective is to ascertain hardware friendly configurations based on the experience span for a limited set of configurations derived from a combination of variables in terms of accuracy, power, and latency. Once the regression fits are established for the metrics of interest, single or multi-objective contours related to these metrics can be created to analytically and experimentally identify the near-optimal solution for deployment. The dissertation also demonstrates that introducing LIME (Local Model Agnostic Explanations)-based sample weights into the training framework or employing metric learning in conjunction with the Reg-Tune method can enhance the inference accuracy of deployed models without altering network parameters. Furthermore, the LIME-based weighting method is shown to be effective for handling unbalanced datasets and mitigating catastrophic forgetting in incremental learning. Additionally, different metrics are profiled across two different device platforms and for a variety of applications, thereby increasing the applicability of the method. For instance, deployments based on the fine-tuning method for physical activity recognition on FPGA demonstrate at least 5.7x better energy efficiency than recent implementations without compromising accuracy. On the other hand, when combined with weighted loss-based training, the Reg-Tune approach identifies a deployment configuration with approximately 8x better energy efficiency than recent keyword spotting implementations on FPGA. Finally, the integration of metric learning with Reg-Tune results in a 14.5x and 101.8x reduction in model size, alongside 2.5x and 5.9x improvements in inference efficiency for keyword spotting and image classification on the Nvidia Jetson Nano 4GB SDK, compared to baseline and recent implementations, respectively.
Protocols for Air-water Communication and Underwater Localization using Nonlinear Optoacoustic Links
(2024-01-01) Mahmud, Muntasir; Younis, Mohamed; Computer Science and Electrical Engineering; Engineering, Electrical
While numerous studies have been dedicated to tackling the challenges of underwater communication and localization, little attention has been paid to effectively establishing direct communication links between aerial and deep underwater nodes (UWN). This dissertation exploits the optoacoustic effect to devise a new solution for directly reaching and localizing underwater nodes from the air. The contribution mitigates the shortcomings of conventional methods which require the involvement of a surface gateway with dual modems to establish radio and acoustic links in the air and underwater, respectively. Deploying such a gateway introduces logistical challenges and security vulnerabilities. Unlike radio, acoustic, and visual light, optoacoustic links involve two distinct signal types, optical (laser beam) in air and acoustic in water. Therefore, the laser beam needs to be modulated such that the resulting acoustic signals can be accurately demodulated to retrieve the data. Such a communication technique is unique. To address challenges in optoacoustic communication, we propose two modulation techniques: Vapor Cloud Delayed-Differential Pulse Position Modulation (VCD-DPPM) and Optical Focusing-based Adaptive Modulation (OFAM). OFAM includes both stationary (1D) and dynamic (3D) focusing for a single laser transmitter, providing stable signal generation even with vapor cloud buildup. The validation results show that VCD-DPPM, OFAM-1D, and OFAM-3D achieve data rates approximately 5.12, 6, and 4.4 times higher than on-off keying (OOK). These techniques also improve power efficiency by 137% over OOK. Machine learning techniques have also been leveraged in the demodulation process for increased robustness, where 94.75% demodulation accuracy is achieved with the Random Forest model. To handle multipath interference in high data rate applications, we have developed a deep learning approach using U-Net for signal equalization and ResNet for symbol detection, achieving 96.6% and 91.7% accuracy, compared to 72.94% and 65.30% with traditional peak detection. The proposed air-water wireless communication protocols are further leveraged to extend the GPS service to the underwater environment by remotely localizing UWN. In our approach, GPS coordinates are transmitted from the air to the UWN via creating an underwater temporary acoustic transmitter (plasma) through the optoacoustic process. We analyze the process of controlling the shape and size of the plasma to control the acoustic signal duration and directivity. First, we have developed a Receive Signal Strength (RSS) based localization method using shorter-more spherical shaped plasma. However, the generated acoustic signal shows more directivity as the plasma shape elongates to achieve higher localization range. Therefore, we have utilized the directive nature of the signal and developed a fully connected deep neural network (DNN) based localization method by determining the receiver angle relative to plasma. The simulation results with laboratory constructed dataset show the effectiveness of our approach. Both of our approaches achieve better accuracy compared to traditional techniques without using surface or underwater anchor nodes.
Building an Efficient PDF Malware Detection System
(2024-01-01) Liu, Ran; Nicholas, Charles; Computer Science and Electrical Engineering; Computer Science
With the widespread use of the Portable Document Format (PDF), it has increasingly becoming a target for malware, highlighting the need for effective detection solutions. In recent years, machine learning-based methods for PDF malware detection have grown in popularity. However, the effectiveness of ML models is closely related to the quality of the training dataset and the employed feature set. Besides, many well-known ML-based detectors require a large number of specimen features to be collected before making a decision, which can be time-consuming. This thesis addresses these challenges by proposing a two-stage approach for PDF malware detection. The initial phase focuses on rapid detection, providing an early warning against potential PDF malware attacks. Following this fast detection stage, we introduce a robust and highly reliable feature set for PDF malware identification. Our contributions are as follows: 1. Rapid Detection Methodology: Compared to traditional machine learning or neural network models, our novel, distance-based method for rapid PDF malware detection requires much less training samples. Evaluated on the Contagio dataset, our method shows that it can detect 90.50\% of malware samples using only 20 benign PDFs for model training.2. Dataset Analysis and Introduction of PdfRep: Through an examination of two widely used PDF malware datasets, namely Contagio and CIC, we find biases and representativeness issues that compromise malware detection model reliability. To mitigate these limitations, we present PdfRep, a new, more representative PDF malware dataset that outperforms existing PDF malware dataset in evaluation metrics. 3. Compact Feature Set for Enhanced Robustness: We introduce a novel set of just five features designed to optimize training efficiency and enhance the robustness of PDF malware detection systems. Experiments show that this compact feature set strengthens PDF malware detection systems against particular adversarial attacks and allows the building of highly accurate models. In summary, this thesis presents a comprehensive approach for PDF malware detection, from rapid initial alerts to the deployment of a robust, efficient detection system enhanced by a novel dataset and an optimized feature set. When taken as a whole, these developments offer significant progress to defend against PDF-based malware attacks.
Flexible FMRI Data Analysis Using Machine Learning
(2024-01-01) Jin, Rui; Kim, Seung-Jun; Computer Science and Electrical Engineering; Engineering, Electrical
Functional magnetic resonance imaging (FMRI) is a non-invasive neuroimaging tool for capturing brain neural activations. This dissertation proposes various machine learning methods for fMRI data analysis focusing on estimating the neural activation maps and analyzing their relations to brain functions and disorders. First, a novel dictionary learning (DL) method is proposed for estimating brain neural activation maps by exploiting the sparse nature of brain activations. The subject group attributes are incorporated into a supervised DL formulation to characterize the activation maps that are shared across subject groups and those that can explain the group differences. The proposed method is tested with real task fMRI data sets from schizophrenic subjects and healthy controls. The benefits of our method are reflected on generating more stable results, finding novel maps that are not estimated from benchmark methods, and estimating task-related maps that are able to significantly discriminate the subject groups. The results are further validated using a correlation analysis with neuropsychological test scores. Then, the aforementioned DL method is extended for analyzing multisubject multiset fMRI data sets. The subject group attributes are again incorporated. The multiple fMRI data sets, such as the task fMRI data sets acquired from different tasks, are analyzed jointly. From the analysis, four types of maps emerge. Namely, the maps that are either shared or showing group differences across subject groups are estimated. Also, the maps that are common across data sets and those that are unique to a data set are identified. Importantly, the algorithm can flexibly determine the map types without rigid allocation of them in the factors. Both synthetic and real data experiments show the benefits of the proposed approach. Finally, a deep object-centric learning-based fMRI data analysis method is proposed to estimate the variabilities of brain neural activation maps over fMRI volumes. The matrix factorization-based approaches such as the DL models assume a common subspace for the neural activation maps, thereby limiting the ability of capturing rich variabilities. The proposed method regards each map as an “object” with latent variables learned using an autoencoder with an attention mechanism. Learning efficient representations in the latent space encourages the model to learn a set of consistent maps across the fMRI volumes. The 3D convolutional neural network is utilized as the main building block of the proposed architecture. Experiments using both synthetic and real fMRI data sets verify the advantages of the proposed approach compared to existing matrix factorization-based methods.
QUANTIZED LARGE LANGUAGE MODELS FOR MENTAL HEALTH APPLICATIONS: A BENCHMARK STUDY ON EFFICIENCY, ACCURACY AND RESOURCE ALLOCATION
(2024-01-01) Jannumahanti, Aayush; Gaur, Dr. Manas; Raff, Dr. Edward; Computer Science and Electrical Engineering; Computer Science
Quantization is a technique that compresses numerical representations to reduce space requirements, though it may sacrifice some precision. Albeit this lossy compression method can improve efficiency, it often comes at the cost of performance. Large Language Models (LLMs) are computationally intensive, posing challenges for users with limited hardware resources. However, advancements in fine-tuning strategies such as QLoRa, LLM.int8(), GGUF, GGML, llama.cpp, and various quantization techniques (8/4-bit, NF4, FP16/32/64, BF16, bitsandbytes) have democratized access to LLMs by reducing the resource burden. LLM weights are typically stored as floating-point numbers, and quantization reduces the precision of these weights to decrease the model’s resource requirements. While this can significantly reduce model size, it may also impact accuracy due to the compressed representation of weights. Lower levels of quantization result in smaller models but may lead to diminished performance. The findings from this research provide critical insights into the viability of using quantized LLMs in sensitive domains like mental health. They highlight the importance of balancing explanation quality with computational efficiency. This benchmarking effort lays the groundwork for deploying effective and resource-efficient LLMs in mental health applications, ultimately supporting professionals and patients with reliable AI-driven insights. As the study progresses, models will be trained sequentially in groups, categorized by familiessuch as LLAMA, Phi, Mixtral, Hermes, Falcon, Gemma, Qwen, and others. This research explores the trade-off between weight precision and model accuracy, aiming to better understand the challenges and potential of quantized LLMs in mental health applications. All models were trained and tested with the generous support of the University of Maryland, Baltimore County’s High-Performance Computing Facility, which provided GPU-accelerated resources.

Browse

Recent Submissions