Search

The current seismic data processing pipeline is surprisingly human-dependent. With the rapid increase of seismic-sensor data availability, all manual data processing approaches fail to detect, classify, and analyze seismic activity within a reasonable amount of time. An automated, fast, and reliable seismic data processing pipeline is desired for the meaningful analysis of massive seismic datasets. In this thesis, we show how advanced time-series data-mining and machine learning techniques can be leveraged to resolve this issue. We precisely focus on seismic activity detection, classification, and inspection using our techniques that would help us better understand the surrounding earth structure, earthquake evaluation, and seismic monitoring In this dissertation, (a) we demonstrate a semi-supervised motif discovery algorithm that forms a nearest neighbor graph to discover novel seismic events from static continuous waveforms. (b) We exhibit a seismic data repository system that can extract thousands of seismic waveforms including annotations using complex queries within seconds. (c) We design and implement a hierarchical neural network that can predict seismic depth from seismograms and classify deep and shallow earthquakes with 86.5% F1 score.

The rise of the Internet of Things (IoT) services has facilitated data sharing and information diffusion among billions of devices. However, IoT still faces data privacy and energy consumption challenges yet to be resolved. Here, the data privacy challenge implies that clients in terms of IoT devices have to share their sensitive data (such as clients’ locations) with a centralized server or data center (managed by third parties), which analyze these sensitive data to generate high-level knowledge that can provide better services to the clients. The energy consumption challenge implies that clients have to frequently send their data to a centralized server, thus quickly draining their portable batteries. In this dissertation, we apply Federated Learning (FL) to IoT in order to resolve the data privacy and energy consumption challenges. FL is a distributed machine learning technique to allow clients in terms of IoT devices to locally train the models/knowledge. Instead of sharing raw data sets, the clients in FL only share their derived local models with the FL server, which aggregates the local models from the clients to generate a global model that can better fit the data sets of the clients. FL does not require clients to share their raw data sets, and so it can potentially address data privacy issues and reduce the energy consumption of clients since massive data transmissions are not necessary for training models. Due to the heterogeneous feature of the clients in wireless FL, where different clients have different computing and communications capabilities, client selection is critical to the performance of FL in terms of model training time and model accuracy. The goal of client selection is to maximize the number of qualified clients, who can successfully upload their local models to the FL server via the base station (BS) before the predefined deadline, in each global iteration. Client selection also depends on the wireless access technology applied at the

Spatially resolved transcriptomics technologies investigate gene expression in the context of a tissue. While imaging-based technologies enable gene expression quantification at a single-cell level, only two NGS-based technologies are currently capable of single-cell resolution. Both Stereo-seq and Seq-Scope involve two rounds of sequencing: one to sequence the location of spatial barcodes and the other to obtain sequence of cDNA. Thus, the objective of the research study is to develop a technology that can capture transcripts from every cell of a tissue without the need for two rounds of sequencing. To this end, a capture surface or barcoded array (similar to 10x Genomics’ Visium platform) with feature sizes of 2.0µm and 0.6µm was manufactured. Spatial transcriptomics was performed on a fresh-frozen and FFPE mouse brain tissue section. The research study demonstrates successful capture of transcripts from the tissue. Furthermore, spatial domains and spatially variable genes identified were consistent with published literature.

In 1993, Dwork and Naor proposed using computational puzzles, a resource burning mechanism, to combat spam email. In the ensuing three decades, resource burning has broadened to include communication capacity, computer memory, and human effort. It has become a well-established tool in distributed security. Due to the cost attached to utilizing resource burning mechanism, these have not been popularized in domains apart from cryptocurrency. In this dissertation, we design efficient resource burning based Sybil defense techniques for permissionless systems. As a first step, we identify existing resource burning mechanisms in literature in Chapter 2. Additionally, we enumerate numerous open problems in a wide variety of distributed systems, where resource burning might come in handy. Next, we develop scalable distributed protocols to reduce the fraction of Sybil participants in large scale permissionless systems. To this end, we design a family of provably secure resource burning protocols that have a resource cost which is commensurate with that of the attacker and the number of honest participants that join the system in Chapter 3. In practice, we observed that the behavior of the join and departure of non-malicious IDs does not vary rapidly over a short duration of time. Using this insight, we design an algorithm that estimates the join rate of non-malicious IDs in the system in Chapter 4. Later, we use this algorithm to determine if the system is under attack. Our third contribution in this dissertation is a Sybil defense protocol that guarantees an algorithmic spend rate that is a slow-growing function of the resource cost to the adversary in Chapter 5. Additionally, we ensure that in the absence of an attack, the algorithmic spending is commensurate with the join rate of non-malicious IDs. To test the theory in practice, we simulate the various Sybil defense protocols proposed throughout this dissertation on several real-worl

Existing communication protocols in security networks are highly centralized. While this naively makes the controls easier to physically secure, external actors require fewer resources to disrupt the system because there are fewer points in the system can be interrupted without the entire system failing. We present a solution to this problem using a proof-of-work-based blockchain implementation built on MultiChain. We construct a test-bed network containing visual imagers and microwave sensor information. These data types are ubiquitous in perimeter security systems and allow a realistic representation of a real-world network architecture. The cameras in this system use an object detection algorithm to find important targets in the scene. The raw data from both the sensors and imagers are placed in a transaction. These transactions are then bundled into blocks and broadcast to the rest of the network using the Bitcoin-based MultiChain protocol. We develop five tests to examine the security metrics of our network. We performed the five security metric test using different sized networks from 7 to 39 nodes to determine how the metrics scale with respect to size. We find that when compared to a centralized architecture our implementation provides a resiliency increase that is expected from a blockchain- based protocol without slowing the system so much that a human operator would notice. Furthermore, our approach is able to detect tampering in real time. Based on these results, we theorize that security networks in general could use a blockchain- based approach in a meaningful way.

In 1993, Dwork and Naor proposed using computational puzzles, a resource burning mechanism, to combat spam email. In the ensuing three decades, resource burning has broadened to include communication capacity, computer memory, and human effort. It has become a well-established tool in distributed security. Due to the cost attached to utilizing resource burning mechanism, these have not been popularized in domains apart from cryptocurrency. In this dissertation, we design efficient resource burning based Sybil defense techniques for permissionless systems. As a first step, we identify existing resource burning mechanisms in literature in Chapter 2. Additionally, we enumerate numerous open problems in a wide variety of distributed systems, where resource burning might come in handy. Next, we develop scalable distributed protocols to reduce the fraction of Sybil participants in large scale permissionless systems. To this end, we design a family of provably secure resource burning protocols that have a resource cost which is commensurate with that of the attacker and the number of honest participants that join the system in Chapter 3. In practice, we observed that the behavior of the join and departure of non-malicious IDs does not vary rapidly over a short duration of time. Using this insight, we design an algorithm that estimates the join rate of non-malicious IDs in the system in Chapter 4. Later, we use this algorithm to determine if the system is under attack. Our third contribution in this dissertation is a Sybil defense protocol that guarantees an algorithmic spend rate that is a slow-growing function of the resource cost to the adversary in Chapter 5. Additionally, we ensure that in the absence of an attack, the algorithmic spending is commensurate with the join rate of non-malicious IDs. To test the theory in practice, we simulate the various Sybil defense protocols proposed throughout this dissertation on several real-worl

Existing communication protocols in security networks are highly centralized. While this naively makes the controls easier to physically secure, external actors require fewer resources to disrupt the system because there are fewer points in the system can be interrupted without the entire system failing. We present a solution to this problem using a proof-of-work-based blockchain implementation built on MultiChain. We construct a test-bed network containing visual imagers and microwave sensor information. These data types are ubiquitous in perimeter security systems and allow a realistic representation of a real-world network architecture. The cameras in this system use an object detection algorithm to find important targets in the scene. The raw data from both the sensors and imagers are placed in a transaction. These transactions are then bundled into blocks and broadcast to the rest of the network using the Bitcoin-based MultiChain protocol. We develop five tests to examine the security metrics of our network. We performed the five security metric test using different sized networks from 7 to 39 nodes to determine how the metrics scale with respect to size. We find that when compared to a centralized architecture our implementation provides a resiliency increase that is expected from a blockchain- based protocol without slowing the system so much that a human operator would notice. Furthermore, our approach is able to detect tampering in real time. Based on these results, we theorize that security networks in general could use a blockchain- based approach in a meaningful way.

In a time when data is being constantly generated by phones, vehicles, sensor net- works, social media, etc. detecting anomalies with in the data can be very crucial. In cases where we know little prior knowledge about the data, it becomes difficult to extract uncertainty about our results. In this thesis, we will propose a framework in which we can extract uncertainty distributions from data-driven modeling prob- lems. We will show some concrete examples of how to apply framework and provide some insight into what the uncertainty distributions are telling us using High Density Regions (HDRs).

Simulation of large molecular structures and their interactions has become a major component of modern biomolecular research. Methods to simulate these type of molecules span a wide array of resolutions, from all atom molecular dynamics to model interaction energetics to systems of linear equations to evaluate population kinetics. In recent years, there has been an acceleration of molecular structural information production, primarily from x-ray crystallography and electron microscopy. This data has provided modelers the ability to produce better representations of these molecular structures. The purpose of this research is to take advantage of this information to develop multi-resolution models for the analysis of large molecule motions and interactions. Our methodology focuses on the use of structural models of a given biological system and simulating the molecules using different conditions (number or ratio of molecules being simulated) and constraints (rigid or semi-flexible models). We combine computational geometry and statistical techniques to perform efficient structural modeling and simulation. Our goal is to utilize our methods to analyze the effect of geometry on molecule interactions, e.g., shape of packed structures or influences of steric hindrance caused by interacting molecules. We focus our work on larger molecular systems, both in size of the molecular structures and number of interacting molecules. The focus of our evaluation is the human allergic immune response. The immune response is triggered by cell surface molecular aggregation of antibodies via an antigen. With our analysis we gain insight into how different allergen geometries affect the size and shape of aggregate structures that form on the cell surface. We perform a multi-resolution analysis of our structures and model the problem in two ways, a lower resolution rigid body representation which can model the aggregation process, and a higher resolution flexible model which can be used

In a time when data is being constantly generated by phones, vehicles, sensor net- works, social media, etc. detecting anomalies with in the data can be very crucial. In cases where we know little prior knowledge about the data, it becomes difficult to extract uncertainty about our results. In this thesis, we will propose a framework in which we can extract uncertainty distributions from data-driven modeling prob- lems. We will show some concrete examples of how to apply framework and provide some insight into what the uncertainty distributions are telling us using High Density Regions (HDRs).

Simulation of large molecular structures and their interactions has become a major component of modern biomolecular research. Methods to simulate these type of molecules span a wide array of resolutions, from all atom molecular dynamics to model interaction energetics to systems of linear equations to evaluate population kinetics. In recent years, there has been an acceleration of molecular structural information production, primarily from x-ray crystallography and electron microscopy. This data has provided modelers the ability to produce better representations of these molecular structures. The purpose of this research is to take advantage of this information to develop multi-resolution models for the analysis of large molecule motions and interactions. Our methodology focuses on the use of structural models of a given biological system and simulating the molecules using different conditions (number or ratio of molecules being simulated) and constraints (rigid or semi-flexible models). We combine computational geometry and statistical techniques to perform efficient structural modeling and simulation. Our goal is to utilize our methods to analyze the effect of geometry on molecule interactions, e.g., shape of packed structures or influences of steric hindrance caused by interacting molecules. We focus our work on larger molecular systems, both in size of the molecular structures and number of interacting molecules. The focus of our evaluation is the human allergic immune response. The immune response is triggered by cell surface molecular aggregation of antibodies via an antigen. With our analysis we gain insight into how different allergen geometries affect the size and shape of aggregate structures that form on the cell surface. We perform a multi-resolution analysis of our structures and model the problem in two ways, a lower resolution rigid body representation which can model the aggregation process, and a higher resolution flexible model which can be used

Due to an increase in regulatory scrutiny and medical facility accreditation requirements to monitor patient radiation dose from diagnostic imaging procedures, there is a growing necessity to determine and record accurate patient radiation dose from diagnostic medical imaging procedures. Current methods of patient dosimetry in diagnostic imaging are both extremely difficult and time consuming, require large computing resources (such as Monte Carlo computations), or lack accuracy due to using data based on homogeneous materials and “standard-man sized” anthropomorphic models. This dissertation provides an algorithm that calculates a more accurate dose using patient-specific projection radiographic images. The algorithm includes measurements acquired during routine physics testing of the x-ray unit, data from two-view (Anterior-Posterior and Lateral chest) radiographic images, and accounts for patient specific and body habitus. The algorithm developed in this work uses formulas for calculating entrance and exit dose utilizing a new dose correction factor. The dose correction factor is based on the exposure index and average grayscale from radiographic images specific to the subject and shows improved accuracy of traditional calculation methods for entrance and exit dose calculations. Measurements using optically stimulated luminescent dosimeters, physics measurements on radiographic equipment, and Monte Carlo simulations were used to test and develop the algorithm. This new algorithm uses the mean radiographic image grayscale value over the region of interest and the reported Exposure Index to create a correction factor to correct patient dose calculations. The final product is an algorithm for calculating patient specific dose from AP and lateral chest radiographs that is more accurate with less associated uncertainty than current traditional dose calculation methods.

The Bragg peak geometry of the depth dose distributions for hadrons allows for precise and effective dose delivery to tumors while sparing neighboring healthy tissue. Further, compared against other forms of radiotherapeutic treatments, such as electron beam therapy (EBT) or photons (x and \(\gamma\)-rays), hadrons create denser ionization events along the particle track, which induces irreparable damage to DNA, and thus are more effective at inactivating cancerous cells. The measurement of radiation's ability to inactivate cellular reproduction is the relative biological effectiveness (RBE). A quality related to the RBE that is a measurable physical property is the linear energy transfer (LET), a measurement of the deposited energy over the particle track. The purpose of this work was to develop a method of measuring LET of radiotherapeutic beams using an air-filled ionization chamber (IC) and liquid-filled ionization chamber (LIC). Through analysis of the IC and LIC signals, we determined a correlation between the recombination index (IC/LIC) and dose-averaged LET.

In the era of new technologies, computer scientists deal with massive data of size hundreds of terabytes. Smart cities, social networks, health care systems, large sensor networks, etc. are constantly generating new data. It is non-trivial to extract knowledge from big datasets because traditional data mining algorithms run impractically on such big datasets. However, distributed systems have come to aid this problem while introducing new challenges in designing scalable algorithms. The transition from traditional algorithms to the ones that can be run on a distributed platform should be done carefully. Researchers should design the modern distributed algorithms based on the problem domain. The main goal of this dissertation is to demonstrate the importance of domain specific knowledge in developing scalable knowledge discovery algorithms on distributed systems. Data properties such as origin, type, context and size play important roles to achieve speed, efficiency and scalability. In this dissertation, I describe three domain specific knowledge discovery systems on three diverse domains: a distributed algorithm to extract patterns from log messages generated by computers, a distributed algorithm to find abnormal behavior in social media, and a scalable algorithm for matching patterns in streaming time series data. I explain how to exploit the data properties in a distributed knowledge discovery system to achieve scalability and speed. The algorithms achieve horizontal scalability for any data size, and the systems are currently deployed at the University of New Mexico.

Online data contains a wealth of information, but as with most user-generated content, it is full of noise, fraud, and automated behavior. The prevalence of "junk" and fraudulent text affects users, businesses, and researchers alike. To make matters worse, there is a lack of ground truth data for these types of text, and the appearance of the text is constantly changing as fraudsters adapt to pressures from hosting sites. The goal of my dissertation is therefore to extract high-quality content from and identify fraudulent and automated behavior in large, complex social media datasets in the absence of ground truth data. Specifically, in my dissertation I design a collection of data inspection, filtering, fusion, mining, and exploration algorithms to: automate data cleaning to produce usable data for mining algorithms, quantify the trustworthiness of business behavior in online e-commerce sites, and efficiently identify automated accounts in large and constantly changing social networks. The main components of this work include: noise removal, data fusion, multi-source feature generation, network exploration, and anomaly detection.

Online data contains a wealth of information, but as with most user-generated content, it is full of noise, fraud, and automated behavior. The prevalence of "junk" and fraudulent text affects users, businesses, and researchers alike. To make matters worse, there is a lack of ground truth data for these types of text, and the appearance of the text is constantly changing as fraudsters adapt to pressures from hosting sites. The goal of my dissertation is therefore to extract high-quality content from and identify fraudulent and automated behavior in large, complex social media datasets in the absence of ground truth data. Specifically, in my dissertation I design a collection of data inspection, filtering, fusion, mining, and exploration algorithms to: automate data cleaning to produce usable data for mining algorithms, quantify the trustworthiness of business behavior in online e-commerce sites, and efficiently identify automated accounts in large and constantly changing social networks. The main components of this work include: noise removal, data fusion, multi-source feature generation, network exploration, and anomaly detection.

In the era of new technologies, computer scientists deal with massive data of size hundreds of terabytes. Smart cities, social networks, health care systems, large sensor networks, etc. are constantly generating new data. It is non-trivial to extract knowledge from big datasets because traditional data mining algorithms run impractically on such big datasets. However, distributed systems have come to aid this problem while introducing new challenges in designing scalable algorithms. The transition from traditional algorithms to the ones that can be run on a distributed platform should be done carefully. Researchers should design the modern distributed algorithms based on the problem domain. The main goal of this dissertation is to demonstrate the importance of domain specific knowledge in developing scalable knowledge discovery algorithms on distributed systems. Data properties such as origin, type, context and size play important roles to achieve speed, efficiency and scalability. In this dissertation, I describe three domain specific knowledge discovery systems on three diverse domains: a distributed algorithm to extract patterns from log messages generated by computers, a distributed algorithm to find abnormal behavior in social media, and a scalable algorithm for matching patterns in streaming time series data. I explain how to exploit the data properties in a distributed knowledge discovery system to achieve scalability and speed. The algorithms achieve horizontal scalability for any data size, and the systems are currently deployed at the University of New Mexico.

The Bragg peak geometry of the depth dose distributions for hadrons allows for precise and effective dose delivery to tumors while sparing neighboring healthy tissue. Further, compared against other forms of radiotherapeutic treatments, such as electron beam therapy (EBT) or photons (x and \(\gamma\)-rays), hadrons create denser ionization events along the particle track, which induces irreparable damage to DNA, and thus are more effective at inactivating cancerous cells. The measurement of radiation's ability to inactivate cellular reproduction is the relative biological effectiveness (RBE). A quality related to the RBE that is a measurable physical property is the linear energy transfer (LET), a measurement of the deposited energy over the particle track. The purpose of this work was to develop a method of measuring LET of radiotherapeutic beams using an air-filled ionization chamber (IC) and liquid-filled ionization chamber (LIC). Through analysis of the IC and LIC signals, we determined a correlation between the recombination index (IC/LIC) and dose-averaged LET.

As the resolution of current image sensors is increasing and the readout electronics is getting faster, the amount of data produced by these imagers is becoming excessively large to store, transmit, and analyze. As a result, sparse data representation has become an interesting research topic in the last few years. Human eye contains millions of photoreceptors, however, only some sparse data is transmitted to the optic nerves. Image processing techniques implemented on a chip and at the pixel level is the main building blocks for retina-like sensors. In this way, instead of transmitting raw images that require massive storage, and off-line processing, the imager transmits only vital information that is relevant to the application of interest. Inspired by the human eye and the way retina handles the data, we have implemented two different readout integrated circuits. In the first method, we have modified a conventional CTIA unit cell to implement in-pixel multispectral classification in the analog domain. The ROIC is designed to utilize spectrally tunable dot-in-a-well (DWELL) infrared photodetector to exploit the possibility of real-time on-chip multispectral imaging for classification. The unit cells are designed to include all necessary elements needed for spectral classification, including the support for high-voltage time varying positive and negative biases, bipolar integration, and selective sample-and-hold circuits. A test chip was designed and fabricated using TSMC's 350nm high voltage CMOS process (CL035HV-DDD) technology. Comprehensive pre-silicon verification proved functionalities of the chip. A custom reconfigurable PCB board and a flexible FPGA firmware are developed to test the chip in cryostat condition. Initial testing results validate the design specs. In the second method, we report a readout integrated circuit featured with a fine control over the bias of individual pixels in every frame independently. To exploit multispectral spatiotemporal im

Selecting an appropriate prior distribution is a fundamental issue in Bayesian Statistics. In this dissertation, under the framework provided by Berger and Bernardo, I derive the reference priors for several models which include: Analysis of Variance (ANOVA)/Analysis of Covariance (ANCOVA) models with a categorical variable under common ordering constraints, the conditionally autoregressive (CAR) models and the simultaneous autoregressive (SAR) models with a spatial autoregression parameter considered. The performances of reference priors for ANOVA/ANCOVA models are evaluated by simulation studies with comparisons to Jeffreys' prior and Least Squares Estimation (LSE). The priors are then illustrated in a Bayesian model of the "Risk of Type 2 Diabetes in New Mexico" data, where the relationship between the type 2 diabetes risk (through Hemoglobin A1c) and different smoking levels is investigated. In both simulation studies and real data set modeling, the reference priors that incorporate internal order information show good performances and can be used as default priors. The reference priors for the CAR and SAR models are also illustrated in the "1999 SAT State Average Verbal Scores" data with a comparison to a Uniform prior distribution. Due to the complexity of the reference priors for both CAR and SAR models, only a portion (12 states in the Midwest) of the original data set is considered. The reference priors can give a different marginal posterior distribution compared to a Uniform prior, which provides an alternative for prior specifications for areal data in Spatial statistics.

Selecting an appropriate prior distribution is a fundamental issue in Bayesian Statistics. In this dissertation, under the framework provided by Berger and Bernardo, I derive the reference priors for several models which include: Analysis of Variance (ANOVA)/Analysis of Covariance (ANCOVA) models with a categorical variable under common ordering constraints, the conditionally autoregressive (CAR) models and the simultaneous autoregressive (SAR) models with a spatial autoregression parameter considered. The performances of reference priors for ANOVA/ANCOVA models are evaluated by simulation studies with comparisons to Jeffreys' prior and Least Squares Estimation (LSE). The priors are then illustrated in a Bayesian model of the "Risk of Type 2 Diabetes in New Mexico" data, where the relationship between the type 2 diabetes risk (through Hemoglobin A1c) and different smoking levels is investigated. In both simulation studies and real data set modeling, the reference priors that incorporate internal order information show good performances and can be used as default priors. The reference priors for the CAR and SAR models are also illustrated in the "1999 SAT State Average Verbal Scores" data with a comparison to a Uniform prior distribution. Due to the complexity of the reference priors for both CAR and SAR models, only a portion (12 states in the Midwest) of the original data set is considered. The reference priors can give a different marginal posterior distribution compared to a Uniform prior, which provides an alternative for prior specifications for areal data in Spatial statistics.