College of Engineering and Computing

Permanent URI for this collection

https://hdl.handle.net/1920/2847

Browse

Now showing 1 - 20 of 694

Driver Behavior Analysis under Naturalistic Driving Scenarios
(2023-11-29) Adhikari, Bikram; Duric, Zoran
Driver behavioral abnormalities account for over 70% of traffic fatalities in the United States, making it the most prominent cause. With the advancement of intelligent transportation systems and the integration of mixed traffic with both automated and driverdriven vehicles, analyzing the behavior of drivers is crucial for ensuring better safety for both drivers and the overall traffic interaction. However, current research is limited to investigating driver behavior in simulated and controlled experiments, often exploring a single feature-set for behavioral analysis. Moreover, research has been constrained in reporting behavioral changes with additional variability in driving experience, time of day, and traffic conditions. In this thesis, we present a comprehensive comparative study between simulated and real-world controlled and naturalistic driving experiments, expanding on publicly available datasets and our own experiments in both simulation and real-world scenarios. The thesis further delves into the comparison between popular modules of data collection, elaborating on the trade-off between cost efficiency and data quality. Finally, the thesis introduces a naturalistic driving dataset consisting of driver behavior captured using driver-centric and vehicle-centric visual and non-visual features. Using the dataset and the collected data, the thesis explores changes in driver behavior categorized based on the experiment’s time of day and driver experience. The research and data presented aim to assist further studies on driver behavior in actual traffic conditions and contribute to the development of better Advanced Driver Assistance Systems (ADAS) for improved driver and road safety.
Miniature Autonomous Robotic Fulfillment System
(2023-11-30) Duric, Zoran
In this paper, we created an automated miniature fulfillment system using robotic manipulators, mobile robots, a rotatory and a box sorter that moves and sorts boxes. The robotic manipulators use computer vision to detect and grasp the box and is robust to changes in position and orientation. The mobile robots uses computer vision for navigation, making it robust to drifting. The robots are all powered by their own Raspberry Pi and share information to each other and the network. We have also connected IR occupancy sensors overhead that can communicate with the robot if there are humans nearby. The objective of this system is to create a testbed for cyberphysical system security, computer vision and inter-robot communication.
Temperature Effect on DC-DC Resonant Converter
(2023-11-27) Li, Qiliang; Huang, Liling
Wide bandgap (WBG) semiconductor devices have been extensively studied for applications in many fields, such as automotive, renewable energy and communication sectors. Silicon carbide (SiC) Metal-Oxide-Semiconductor Field Effect Transistors (MOSFETs) exhibit great performance in modern power converters due to their high thermal conductivity, large breakdown voltage and fast saturation drift velocity. Many topologies of photovoltaic applications can obtain high conversion efficiency using SiC power MOSFET due to the superior properties of fast switching, low conduction loss and high thermal conductivity when compared to Silicon counterparts. The integration of resonant tank on the secondary side of the transformer leads to a high peak efficiency of over 96%. In this thesis, an extensive comparative analysis between Silicon (Si) and Silicon Carbide (SiC) MOSFET was conducted in a 300W full bridge DC-DC resonant converter with a boosting cell rectifier. The analysis encompassed different ambient temperatures and load resistances through gamma-ray irradiation for photovoltaic system with a wide input voltage range. A PSIM thermal model was employed to conduct power loss analysis in a specific Photovoltaic (PV) system involving the variation of On-State Resistance (𝑅𝑅𝑂𝑂𝑂𝑂) across different factors such as temperature and gamma irradiation. In the simulation result, as the junction temperatures increases, SiC devices exhibit lower total power losses than Si devices.
A Systematic and Comparative Study in Deep Learning Approaches in Automated Extraocular Muscle Segmentation and Analysis in Orbit Magnetic Resonance Images
(2023-08-08) Qureshi, Amad Aamir; Wei, Qi
Strabismus is an ocular condition characterized by binocular misalignment, which impacts about 5% of the global population. It can cause double vision, reduced vision, and impair the quality of life. Accurate diagnosis and treatment planning often benefits from the anatomical evaluation of the extraocular muscles (EOMs) that can be obtained by imaging modalities, such as magnetic resonance imaging (MRI). Such image-based examination requires segmenting the ocular structures from images, which is a labor and time-intensive task, subject to error when done manually. Deep learning-based segmentation has shown promise to outline anatomical structures automatically and objectively. We performed three sets of experimentation for EOM segmentation via DL-methods. Furthermore, we analyzed the performance of the deep learning methods through F-measure-based metrics, intersection over union (IoU) and Dice coefficient, and estimation of the EOM centroid (centroid offset). We first investigated the performance of U-Net, U-NeXt, DeepLabV3+, and ConResNet in multi-class pixel-based segmentation of the EOMs from ocular MRI taken in the quasi-coronal plane. Based on the performance evaluation (visual and the quantitative metrics mentioned), the U-Net model achieved the highest overall segmentation accuracy, and lowest centroid offset. It was noted that segmentation accuracy varied in spatially different image plane – relative to the middle slice (optic nerve junction point) in the MRI stack. In the second set of experiments, we compared the performance of the U-Net model with its variants, U-NeXt, Attention U-Net and FD-UNet and subjected the prediction outputs to the same evaluation as before, with U-Net achieving the best performance. We also explored methods in an attempt to improve the model performance – particularly with data augmentation and enhancement, where methods such as Adaptive Gamma Correction and CLAHE enhancement were used with the U-Net model. No significant difference was observed when CLAHE, Adaptive Gamma Correction and a dataset with unenhanced, CLAHE, and adaptive gamma corrected images were tested against unenhanced data, however, did result in better quantitative performance than the standard augmentation technique. Our study provides the insights into the factors that impact the accuracy of deep learning models in segmenting the EOMs, such as spatial slice location, image quality, and contrast and demonstrate the potential of these models in translating into 3D space for potential diagnosis and treatment planning for patients with strabismus and other ocular conditions.
Doping Effects on 2D TMDs and Monolayer FETs with PN-Junction or Heterojunction Channels
(2023) Benyan Eshun, Kwesi; Li, QIliang
The major challenges in the scaling of metal-oxide-semiconductor field effect transistors (MOSFETs) include large source-to-drain leakage and small on/off current ratio, especially when the channel length is below 10nm. As the devices are being shrunk into nanoscale, it is increasingly difficult, if not impossible, to precisely control the dopant position and number in low-dimensional nanomaterials. To address these challenges, we designed and investigated a new type of short-channel FETs based on two-dimensional (2D) transition metal dichalcogenides (TMDCs) with a PN junction or heterojunction at the middle of the channel. The first step is to investigate the doping effects on the electrical properties of n-type and p-type 2D TMDCs, like MoS2 monolayer. In the comprehensive first-principle computational study of 2D TMDC FETs, different gate length and channel materials have been investigated and compared, with a focus on the analysis of ballistic transport, energy band alignment and their impact on channel current density. The results indicate that the off-state leakage current and on/off current ratio are significantly improved in the FETs with a junction channel, in comparison with the FETs with homogeneous channel. Also, the 2DTMDC semiconductors should be protected from unintentional or intentional doping if they are used in the transistors in future integrated circuits. This new junction-channel approach, leveraged with the intrinsic advantages offered by 2D TMDC monolayers, suggests a new and very attractive strategy to construct future nanoelectronic transistors.
Probabilistic Generative Models for Learning with Hypergraphs
(2023) Pedrood, Bahman; Domeniconi, Carlotta; Laskey, Kathryn B
Hypergraphs extend conventional graphs by allowing hyperedges to connect any number of nodes. Despite increasing interest in hypergraphs and numerous attempts to exploit their higher-order connections, existing hypergraph analytics solutions have been limited in their effectiveness, primarily due to the complexity arising from the exponential hyperedge space. This dissertation first examines the efficacy of previous hypergraph analytics methods by developing two network anomaly detection approaches: one based on regular graphs and the other leveraging existing hypergraph analytics techniques. While experimental results demonstrate the effectiveness of the former approach, the latter proves to be less successful, indicating a demand for more advanced hypergraph-based frameworks. To address this demand, this dissertation introduces Hypergraph Simultaneous Generators (HySGen), a versatile probabilistic hypergraph model that formulates its generative process by conditioning the distribution of hyperedges on the nodes' community affiliations. Distinguishing itself from previous models, HySGen does not alter the nature of the hyperedges or constrain their size, and provides a novel method for detecting overlapping communities by effectively utilizing the higher-order connection information in hypergraphs. To tackle the intractable complexity associated with representing the entire state space of the hyperedges, this dissertation introduces a complexity reduction method that reduces the super-exponential inference time to linear without sacrificing any significant precision. Additionally, an algorithmic solution is presented for a runtime issue arising in situations requiring extremely high precision. To objectively assess detected overlapping communities, this dissertation introduces a novel performance measure that addresses the limitations of existing metrics. Experimental results demonstrate the effectiveness, scalability, and superiority of the introduced models and methods compared to current state-of-the-art techniques. The implementations of these models and methods have been incorporated into a popular open-source network analysis and graph mining library. This dissertation lays the foundation for a new line of future research and establishes a framework for developing innovative and effective hypergraph analytics algorithms, addressing the challenges and complexities associated with this field.
Poisson Mixture Network Tomography
(2023) Coblenz, Joshua; Ephraim, Yariv; Mark, Brian L
Network tomography is the study of the internal characteristics of networks from incomplete measurements. Statistical methods are used to infer the parameters of underlying distributions of network traffic. Consider a network consisting of a number of nodes and links. One goal is to estimate mean traffic rates on pairs of source-destination nodes using measurements of traffic on a number of links. This is in essence the solving of an inverse problem of a system of equations. In general, the system of equations is underdetermined, because the number of links on which traffic is being measured is less than the number of source-destination pairs. The network can be deterministic, where the path from source to destination is known a priori, or random, where the path from source to destination is unknown a priori but defined by probabilities. The field was pioneered by Y. Vardi in 1996. One of the methods Vardi used to estimate mean traffic rates is called moment matching. Here, equations relating the first moments and second central moments of the link traffic to the first moments and second central moments of the source-destination traffic are used to estimate the mean source-destination traffic rates. Throughout much of its history, network traffic flows were assumed to be Poisson random processes. Recently, network tomography has been extended to traffic flows that are not necessarily Poisson random processes. The distributions of traffic flow in a given time interval were allowed to be mixtures of Poisson distributions. This approach significantly generalizes the capabilities of network tomography. The mixture distributions utilized here are overdispersed, with variances larger than the means. Thus, they are more suitable for traffic than the Poisson model. Second-order and third-order moment matching approaches are developed for estimating the mean traffic rate for each source-destination pair using the minimum I-divergence iterative procedure. The performance of the proposed approach is demonstrated for several numerical examples and compared to the EM algorithm, which was derived assuming a Gaussian traffic model.
Understanding Novel Energy Transfer Modalities Using DNA Scaffolds for Precise Sub-Nanometer Positioning
(2023) Chiriboga, Matthew; Medintz, Igor L
A central theme in the field of nanophotonics is the characterization of new materials that display unique light-matter interactions. In practice, this knowledge is asserted through an iterative process of model design, assembly, and characterization. The end result of this process is a set of logical design principles that are utilized to manufacture novel nanomaterials. Of particular interest is controlling the transfer of excited state energy between fluorophores separated by distances of less than 20 nm. At this scale, energy transfer through non-radiative dipole-dipole coupling, categorically denoted as resonance energy transfer (RET), dominates the interaction. In RET mechanisms, the energy transfer efficiency between fluorophores is inversely proportional to their separation distance. For example, the most common RET mechanism, Förster resonance energy transfer (FRET), exhibits a 6th power decay as a function of separation distance. As such, small fluctuations in fluorophore positioning will dramatically affect and can ultimately nullify the expected behavior of the system. Therefore, to develop prototype nanosensors exploiting RET, there exists a demand for a robust yet programmable molecular scaffold. Although there are many functional nanomaterials available, none are as versatile or convenient for fluorophore scaffolding as DNA. By exploiting the inherent chemical properties of DNA, scaffolds can be designed to self-assemble into near-arbitrary 2- and 3-dimensional geometries. Furthermore, the chemical nature of DNA makes it amenable to labeling with various optically active materials such as organic fluorophores, colloidal semiconductor nanocrystals, and noble metal nanoparticles. Herein, we use self-assembled DNA nanostructures as a tool to control the positioning and density of fluorophores in resonance energy transfer cascades. Specifically, we use DNA nanostructures to further describe the limits of multidirectional energy transfer between pseudoisocyanine (PIC) dye aggregates. We then go on to identify a previously unreported method of forming PIC aggregates and demonstrate that the photophysical properties of said aggregates can be tuned by merely modifying the underlying DNA sequence. Next, we pivot to a new class of material, namely 2-dimensional colloidal semiconductor nanoplatelets. We utilize a labeled peptide-PNA-DNA hybrid platform to demonstrate energy transfer with the platelets acting as both resonance energy donors and acceptors. Finally, we address the matter of nanostructure formation quality assurance by reconfiguring a deep convolutional neural network (dCNN) to scour atomic force microscope (AFM) images of DNA nanostructures, identifying fully formed nanostructures, and classifying them by respective geometries.
EVALUATING THE RESILIENCE OF VISION TRANSFORMERS TO MEMBERSHIP INFERENCE ATTACKS
(2023) Muhammad, Talha; Khasawneh, Khaled N
Machine learning has come a long way in recent years. Adoption has begun to spreadand recent economic studies have shown the industry is expected to grow with billions of dollars being invested into the research and development of such technologies, which has been supported by the an economic impact study by The White House and the European Union [1]. Recently, the technology has seen a surge in use cases in domains which normally have not been active users of AI and Machine Learning. Specifically, medicine, finance, and national security/defence has picked up and starting taking advantage of AI and Machine Learning models [2] [3] [4]. The recent advances in Machine Learning architectures such as the transformer architecture has only sped up adoption and spread of the technology [5]. However, with the rapid innovation in the area, research needs to be done around security and privacy, to help build defences against attackers attempting to access sensitive data. Much research exists that has studied various attack methodologies against existing machine learning technologies, however, emerging innovative machine learning architectures have yet to be evaluated against existing, and potentially new types of attacks. In this work, we study the robustness of Vision Transformers (ViTs) to existing Member Inference attacks (MIAs), using a custom built, extendable, and salable framework. Vision Transformers are a novel proposal for adapting the Transformer Architecture to accomplish various vision related tasks such as classification [6]. Membership Inference Attacks are a type of adversarial attack in which the attacker attempts to gather information about the dataset on which a target model is trained on. This can lead to attackers accessing private data which may be sensitive. In our experimentation, we use our novel framework to experimentally explore the Vision Transformers robustness against different Membership Inference Attacks. Our experimental result, using the CIFAR-10 and CIFAR-100 datasets, show that ViTs models could provide a more reliable means of preserving the privacy of training data.
Models and Algorithms for Data-Driven Scheduling
(2023) Fan, Zhengyang; Chang, Kuo-Chu; Ji, Ran
The dissertation examines the problem of machine scheduling in an uncertain environment. To address this challenge, the study employs both traditional operations research (OR) techniques and modern prescriptive analytics methodologies. The traditional OR approaches utilized in the study include stochastic programming and distributionally robust optimization (DRO). Three research works were conducted that focus on modeling and solution perspectives for different types of stochastic scheduling problems: 1. Decomposition Algorithm for Stochastic Parallel Machine Scheduling Problem with Sequence Dependent Setups; 2. Distributionally Robust Single Machine Scheduling Under Wasserstein Ambiguity; 3. Copula-Based Wasserstein Distributionally Robust Optimization with Application in Machine Scheduling and Portfolio Optimization. In addition to the traditional OR techniques, the study also employs modern prescriptive analytics methods that integrate machine learning and decision-making. One research work was conducted that focus on building an integrated system for prediction and optimization: 1. Flight and Maintenance Scheduling for Military Aircraft Fleet: A Data-Driven Approach. Overall, the dissertation aims to explore the effectiveness of both traditional OR and modern prescriptive analytics approaches in handling uncertainty in scheduling problems. The results of this research could have practical implications for industries such as manufacturing, logistics, and transportation, where scheduling plays a critical role in operational efficiency.
Towards Robust and Privacy-aware Time Series Data Mining
(2023) Zhang, Li; Lin, Jessica
With the advancement of sensor technology, large volumes of high-resolution time series are collected in a variety of domains and data mining in large-scale time series has attracted great research attention. While many time series data mining methods in large-scale time series are able to achieve satisfactory performance in a wide variety of tasks in ideal cases, they suffer from performance decreases or even complete failure because of environmental noise, unexpected signals, and potential system failures in many more complex real-world applications. In addition, there have been rising concerns about privacy issues for performing data mining since it is often required to have data access to perform certain data mining tasks on high-resolution time series. However, numerous research efforts have found that long shape-based patterns embedded high-resolution time series could contain sensitive information and be misused by a malicious modeler. However, despite there is a large body of privacy literature, how to perform time series data mining tasks while protecting sensitive patterns is surprisingly seldom explored in privacy-preserving literature. In this dissertation, I investigate the robustness issues in two popular time series data mining tasks, time series anomaly detection and time series chain discovery. To mitigate the challenges in robustness, I introduce three new robust methods: a new time series anomaly detection method that resists unknown background patterns; a new chain discovery method that works for time series with a gap or a corrupted segment; a more general chain discovery method that is more robust to noise and pattern fluctuation in more complex time series. To address the privacy issue, I take the first step to investigate the pattern-level privacy problem in time series data mining tasks. I introduced a privacy-aware scheme that allows time series data to be shared between data owners and a service provider via an intermediary data structure without leaking sensitive information about the owners. My proposed method protects sensitive patterns and their locations and supports various downstream time series data mining tasks while maintaining the performance of these tasks.
Deep Latent Variable Models for Learning Representations of Protein Tertiary Structures
(2023) Alam, Fardina Fathmiul; Shehu, Amarda
The key role that the three-dimensional structure of a protein molecule plays in its function and activities in the cell continues to motivate computational research. In particular, we now know that proteins harness their ability to access different structures to regulate their interactions with other molecules. The research presented in this dissertation leverages the growing momentum in generative AI and contributes increasingly sophisticated deep latent variable models that learn informative representations of protein structures. Rigorous empirical evaluation demonstrates the capabilities of these models in sampling the protein structure space and additionally addressing important protein modeling tasks, linking protein structure and function. The models presented in this dissertation learn directly from experimentally-available structures of different protein molecules and generate physically realistic structures of a target protein, enabling us to expand our in-silico characterization of these ubiquitous molecules beyond the static, single-structure view. This dissertation work advances bioinformatics research in molecular biology.
Analytical Tools for Modeling and Forecasting Global Maritime Cargo Flows under Changing and Uncertain Conditions
(2023) Li, Wenjie; Miller-Hooks, Elise
Global maritime shipping is a backbone of logistic operations for global seaborne trade. These systems are complex, containing hundreds of ports and thousands of shipping routes. Understanding and forecasting the system’s behavior given uncertainty in operational conditions due to weather, climate change impacts, workforce fulfillment levels, new opportunities for crossing the globe, disaster events, and more is important to businesses, shippers, carriers, port stakeholders, regions, and others. This dissertation proposes mathematical models and solution algorithms for estimating and forecasting flows through the maritime system and the system’s performance under changing and uncertain conditions. These conditions arise both from new opportunities, such as the potential to employ Arctic passageways with limited special equipment, and existing and new risks, such as from increased wave heights due to climate change, increased occurrence of disaster events, and concerns about workforce availability. These advancements contribute to an ability to maintain a reliable and resilient global maritime shipping network and the supply chains they support.This dissertation contributes to these modeling and forecasting capabilities in three key areas. Specifically, it presents: (1) A high-fidelity and updatable containerized and bulk cargo shipping network representation with 161 seaports covering 52 countries constructed on publicly available, updatable data sources and mixed-integer linear strategic cargo routing model with combined gradient descent and relax-and-fix decomposition solution methods for estimating global seaborn trade flows. (2) A risk-constrained maritime cargo flow model with exact Benders-branch-and-cut and data-driven Bayesian network solution methodologies for predicting changes in global cargo vessel traffic across a maritime network incorporating Arctic ports and passageways under future reduced-ice shipping scenarios. (3) Quantification of the maritime network’s performance under different disaster events ranging from port-related workforce shortages to natural disasters through a capacity-constrained maritime cargo flow optimization model with workforce level-related scenarios and an Equilibrium Program with Equilibrium Constraints (EPEC) formulation for making coalition-based cross-port investments under natural disaster scenarios, respectively. These analytical tools provide modeling and forecasting capabilities needed to, not only give insights to stakeholders involved in global seaborne trade, but to understand the ramifications of changing trade flows, climate conditions and opportunities on economies, local peoples, the environment, and more.
Planner-Guided Robot Swarms
(2023) Schader, Michael; Luke, Sean
Robot swarms have long been a subject of interest in robotics due to the promise of applying swarm virtues to accomplishing real-world tasks. A swarm is a collection of agents, each with limited capabilities on its own, that produces meaningful emergent behavior, ideally resulting in a system that is scalable, robust, and adaptable. Despite such promise, it has proven difficult to apply robot swarms to complex, multi-step missions. Without central direction, coordination and synchronization are major challenges. In other areas, automated planning is often used in such situations, but planning has typically been seen as antithetical to the distributed nature of a true swarm. To address this challenge, I present planner-guided robot swarms, a novel approach that combines logic-based automated planning with swarms of unspecialized agents in order to orchestrate the collective completion of complicated tasks. I show that the general approach works with an online centralized executive managing the agents’ actions, then enhance it by devolving planning and decision-making to each agent, making a fully-decentralized system. I evaluate such a swarm's performance with varying limitations on communications, and explore ways to mitigate problems caused by agents being out of sync with each other. Finally, I present the results of implementing a planner-guided swarm using real robots engaged in a construction-type scenario. Taken together, these parts form a complete description and exploration of my answer to the challenge of accomplishing complicated tasks with a scalable, robust, and adaptable robot swarm.
Machine Learning Enabled Health Monitoring and\\ Diagnosis of Engineering Systems
(2023) Kamranfar, Parastoo; Shehu, Amarda
System health monitoring provides valuable insight into the condition of an in-service engineering system, potentially enabling planners to make better, data-informed decisions regarding system maintenance and operation safety. An increasing number of sensors are now utilized to provide real-time feedback on in-service structures, underscoring the need to automate system health monitoring. In principle, machine learning (ML) approaches provide us with such an opportunity, but off-the-shelf methods are challenged by often small, largely-unlabeled, and imbalanced sensor data. Because of these unique data characteristics, existing ML-based studies suffer from generalizability and reproducibility. This dissertation addresses a family of exemplary problems in ML-based system health monitoring and puts forth novel ML-based computational methods that can handle the small and imbalanced data regimes. Unsupervised and semi-supervised ML methods that additionally hybridize concepts and techniques from multi-objective optimization, signal processing, multi-instance learning, and more are shown here to allow mapping the states of a selective compliance articulated robot arm from acoustic monitoring data, detecting road conditions from smartphone accelerometer data, and discovering anomalies from time-series data of diverse sensory types. The methodologies presented in this dissertation advance ML-enabled system health monitoring.
Accelerating Polynomial Multiplication for Lattice-based Post-Quantum Cryptography
(2023) Nguyen, Duc Tri; Gaj, Kris
Cryptography plays an important part in today’s security and privacy. With the threat of Quantum Computers, modern cryptography will soon be extended with a new Post- Quantum Cryptography (PQC) standard. The new cryptography standard will bring new challenges: larger keys and certificates, more time-consuming operations, the need for novel and effective side-channel protection, and more.In this work, we focus on one aspect of the challenge – high-speed implementations in hardware and software – and one family of PQC algorithms - lattice-based cryptography. In investigating various approaches to speed up post-quantum cryptography, we observe that Number Theoretic Transform (NTT), Fast Fourier Transform (FFT), and Toom-Cook (TC) are crucial algorithms for the major operation of lattice-based cryptography - polynomial multiplication. This thesis discusses two proposed solutions: compact and efficient Number Theoretic Transform in hardware and high-speed Single Instruction, Multiple Data (SIMD)- based approach to implementing polynomial multiplication in software. In the hardware part, we present a conflict-free memory access for Radix-2×2, a variant of Radix-4 NTT implementation. In addition, our proposed solution addresses the disadvantages of the Radix-4 architecture, which only works for N = 4n. Instead, our design works for any N = 2n, with n ≥ 4. Our design is simple and versatile. It supports three major operations of polynomial mul- tiplication in the same circuit: Forward NTT, point-wise multiplication, and Inverse NTT. Compared to related work, our design uses fewer FPGA resources, such as Digital Signal Processing (DSP) units and Block RAMs (BRAMs). In the software part, we focus on optimized constant-time software implementations of three NIST PQC Key Encapsulation Mechanisms (KEMs) - CRYSTALS-Kyber, NTRU, and Saber - and three Digital Signature schemes - Falcon, SPHINCS+, and XMSS, tar- geting ARMv8 microprocessor cores. All optimized implementations include explicit calls to Advanced Single-Instruction Multiple-Data instructions (a.k.a. NEON instructions). Benchmarking is performed using two platforms: 1) MacBook Air, based on an Apple M1 System on Chip (SoC), including four high-performance ’Firestorm’ ARMv8 cores, running with the frequency of around 3.2 GHz, and 2) Raspberry Pi 4, a single-board computer based on the Broadcom SoC, BCM2711, with four 1.5 GHz 64-bit Cortex-A72 ARMv8 cores. We tailored the Toom-Cook algorithm by selecting optimal settings for NTRU and Saber. We demonstrated the advantages of our signed arithmetic implementation of NTT in Falcon and Kyber compared to the unsigned arithmetic NTT. We also improved the performance of hash-based signatures significantly by using the Crypto Instructions of the Apple M1 processor. Overall, our high-speed NEON implementations achieve considerable speed-up compared to the corresponding reference implementations in C. The obtained algorithm rankings are similar to those reported for the Advanced Vector Extensions 2 (AVX2)-based implementations developed by the respective submission teams, running on Intel and AMD processor cores.
Dynamic modeling for functional data
(2023) Cameron, James; Bagchi, Pramita
The first part of this thesis proposes a measure for heteroscedascity for functional data in the Hilbert space $L^2[0,1]$ equipped with a particular inner product using a distance function on the kernel of the associated covariance function. This results in a closed form expression for the minimal distance of the observed functional data to the nearest homoscedasctic process. We develop an estimator for this measure which, under particular conditions on the observed data, converges to a zero mean normal distribution with a closed form variance if the hypothesis of homoscedascity holds. We propose bootstrap estimation to make the estimating the variance computationally tractable and investigate the accuracy and performance of the proposed methods on numerical simulations and a real data case study. This method is further extended theoretically to the cases of precise hypotheses of heteroscedascity and testing on surrogate variables. The second part of this thesis considers the variable selection problem in a functional logistic regression setting, the proposed methodology being adapting the group lasso penalty. The problem of low incidence rate it considered and handled using a bias reduced maximized incidence approach instead of the classic maximum likelihood estimator. Some theoretical results are established and block co-ordinate gradient descent algorithms are implemented to study the methodology in numerical simulations. As a result of these numerical simulations and the fact that the group lasso is not selection consistent, an adaptive group lasso is studied via numerical simulation for the functional logistic framework.
AI-Enhanced Software Vulnerability and Security Patch Analysis
(2023) Wang, Xinda; Sun, Kun
With the increasing popularity of open-source software (OSS), their embedded vulnerabilities have been widely propagating to downstream software. Although timely applying security patches is the best practice to fix vulnerabilities, OSS users are hard to distinguish and prioritize security patches over tons of non-security patches (i.e., bug fixes, feature updates, etc.). Even worse, software vendors may silently release security patches without providing any explicit advisories. While users are unaware of security patches, attackers can still carefully inspect the patch code changes to exploit unpatched software. Therefore, automatically detecting security patches becomes imperative for software maintenance. In this dissertation, I describe my research efforts to address the above problems. First, I introduce an empirical study that reveals the insecure behavior of software vendors during maintenance and discloses the existence of silent security patches. Second, I present PatchDB, the first large-scale real-world patch dataset, that enables the training of data-hungry AI models for patch detection and facilitates future vulnerability/patch analysis research. An unsupervised method is developed to efficiently collect security patch samples from a huge number of unlabeled GitHub commits. Third, I present GraphSPD, a novel graph learning-based approach for automated security patch detection. By combining rich semantic properties of both pre-patch code and post-patch code in a joint graph structure and adopting a tailored multi-attributed graph convolution network to adapt diverse attributes in a patch graph representation, GraphSPD demonstrates state-of-the-art performance and detects 88 new silent security patches in popular GitHub projects.
Photoacoustic Tomography Image Reconstruction Using Deep Learning
(2023) Hsu, Ko-Tsung; Chitnis, Parag
In photoacoustic tomography, sparse sensing and limited view configurationhave been commonly applied to biomedical imaging due to the spatial constraint of the imaging source. In addition, using a large number of sensors requires sophisticated designs of imaging instruments and expensive computations to process multi-channel data, limiting the use of image reconstruction algorithms for real-time applications. In these scenarios, conventional methods fail to reconstruct high-quality images and result in the reconstructed images being contaminated by severe artifacts stemming from limited view and sparse sampling of the signals acquired by transducers. To overcome these challenges, deep learning methods are introduced. The following manuscripts are organizedxvii into three sections: (1) a comprehensive study of deep learning approaches in photoacoustic tomography (2) Fourier neural operator for solving photoacoustic wave equations (3) incorporating learned physical model into learned iterative reconstruction to speed up the photoacoustic tomography reconstruction. The rationale of the manuscripts is based on our findings. In (1), the results demonstrated that model-based learned iterative reconstruction methods perform the best in photoacoustic image reconstruction compared to the pure data-driven deep learning methods. However, superior reconstruction performance comes at the expense of computational speed due to repeatedly running simulations of physical models in the model-based learned iterative reconstruction pipeline. Therefore, in (2), deep learning-based alternatives to the conventional solver for physical models were explored. Based on the findings in (2), learned physical models were incorporated into the learned iterative reconstruction pipeline in (3) which reduced the computation time originally required by conventional methods. Consequently, the proposed approaches maintained a comparable reconstruction performance to the model-based learned iterative methods and had a 9X reduction in computational speed.
Automated Search, Classification, and Metadata Annotation of Peer-Reviewed Publications for NeuroMorpho.Org
(2023) Maraver, Patricia; Ascoli, Giorgio A
Motivation: The biomedical literature is expanding at ever-increasing rates, and it has become extremely challenging for researchers to keep abreast of new data and discoveries even in their own domains of expertise. For example, NeuroMorpho.Org, a sharing platform for digital reconstructions of neural morphology, must evaluate more than 6000 potentially relevant articles per year to identify data of interest. We introduce a crawler based on periodic full-text searches across publisher web portals that automatically finds and assess the likelihood of a publication to be relevant for the project. Furthermore, it extracts key elements of the metadata such as tracing system, species, cell type, and brain region.Results: Without user interactions, the tool retrieves and stores the bibliographic in- formation (full reference, corresponding email contact, and full-text keyword hits) based on pre-set search logic from a wide range of sources including Elsevier, Springer/Nature, PubMed/PubMedCentral, and Google Scholar. Although different publishing sites require different search configurations, the common interface unifies the process from the user per- spective. Once saved, the tool automatically identifies articles describing digitally recon- structed neural morphologies with high accuracy. In addition, the ability to automatically extract key metadata from neural tracing reduces the risk of errors or inconsistencies in the analysis and interpretation of the data. Conclusions: Since deployment, the tool helped NeuroMorpho.Org more than quintu- ple the yearly volume of processed information. Its processing rate of 900 publications per hour is not only amply sufficient to autonomously track new research, but also allowed the successful evaluation of older publications backlogged due to limited human resources. The number of bio-entities found since launching the tool almost doubled while greatly reducing manual labor. The tool is open source, configurable, and simple to use, making it extensible to other biocuration projects.

Browse

Recent Submissions