Search

Research on graph representation learning (GRL) has made major strides over the past decade, with widespread applications in domains such as e-commerce, personalization, fraud & abuse, life sciences, and social network analysis. Despite its widespread success, fundamental questions on practices employed in modern day GRL have remained unanswered. Unraveling and advancing two such fundamental questions on the practices in modern day GRL forms the overarching theme of my thesis. The first part of my thesis deals with the mathematical foundations of GRL. GRL is used to solve tasks such as node classification, link prediction, clustering, graph classification, and so on, albeit with seemingly different frameworks (e.g. Graph neural networks for node/graph classification, (implicit) matrix factorization for link prediction/ clustering, etc.). The existence of very distinct frameworks for different graph tasks has puzzled researchers and practitioners alike. In my thesis, using group theory, I provide a theoretical blueprint that connects these seemingly different frameworks, bridging methods like matrix factorization and graph neural networks. With this renewed understanding, I then provide guidelines to better realize the full capabilities of these methods in a multitude of tasks. The second part of my thesis deals with cases where modeling real-world objects as a graph is an oversimplified description of the underlying data. Specifically, I look at two such objects (i) modeling hypergraphs (where edges encompass two or more vertices) and (ii) using GRL for predicting protein properties. Towards (i) hypergraphs, I develop a hypergraph neural network which takes advantage of the inherent sparsity of real world hypergraphs, without unduly sacrificing on its ability to distinguish non isomorphic hypergraphs. The designed hypergraph neural network is then leveraged to learn expressive representations of hyperedges for two tasks, namely hyperedge classification and hyperedge expansion. Experiments show that using our network results in improved performance over the current approach of converting the hypergraph into a dyadic graph and using (dyadic) GRL frameworks. Towards (ii) proteins, I introduce the concept of conditional invariances and leverage it to model the inherent flexibility present in proteins. Using conditional invariances, I provide a new framework for GRL which can capture protein-dependent conformations and ensures that all viable conformers of a protein obtain the same representation. Experiments show that endowing existing GRL models with my framework shows noticeable improvements on multiple different protein datasets and tasks.

Large-scale deep learning models have reached previously unattainable performance for various tasks. However, the ever-growing resource consumption of neural networks generates large carbon footprint, brings difficulty for academics to engage in research and stops emerging economies from enjoying growing Artificial Intelligence (AI) benefits. To further scale AI to bring more benefits, two major challenges need to be solved. Firstly, even though large-scale deep learning models achieved remarkable success, their performance is still not satisfactory when fine-tuning with only a handful of examples, thereby hindering widespread adoption in real-world applications where a large scale of labeled data is difficult to obtain. Secondly, current machine learning models are still mainly designed for tasks in closed environments where testing datasets are highly similar to training datasets. When the deployed datasets have distribution shift relative to collected training data, we generally observe degraded performance of developed models. How to build adaptable models becomes another critical challenge. To address those challenges, in this dissertation, we focus on two topics: few-shot learning and domain adaptation, where few-shot learning aims to learn tasks with limited labeled data and domain adaption address the discrepancy between training data and testing data. In Part 1, we show our few-shot learning studies. The proposed few-shot solutions are built upon large-scale language models with evolutionary explorations from improving supervision signals, incorporating unlabeled data and improving few-shot learning abilities with lightweight fine-tuning design to reduce deployment costs. In Part 2, domain adaptation studies are introduced. We develop a progressive series of domain adaption approaches to transfer knowledge across domains efficiently to handle distribution shifts, including capturing common patterns across domains, adaptation with weak supervision and adaption to thousands of domains with limited labeled data and unlabeled data.

The exponential growth of Internet of Things (IoT) and cyber-physical systems is resulting in complex environments comprising of various devices interacting with each other and with users. In addition, the rapid advances in Artificial Intelligence are making those devices able to autonomously modify their behaviors through the use of techniques such as reinforcement learning (RL). There is thus the need for an intelligent monitoring system on the network edge with a global view of the environment to autonomously predict optimal device actions. However, it is clear however that ensuring safety and security in such environments is critical. To this effect, we develop a constrained RL framework for IoT environments that determines optimal devices actions with respect to user-defined goals or required functionalities using deep Q learning. We use anomaly based intrusion detection on the network edge to dynamically generate security and safety policies to constrain the RL agent in the framework. We analyze the balance required between ���safety/security��� and ���functionality��� in IoT environments by manipulating the exploration of safe and unsafe benefit state spaces in the RL framework. We instantiate the framework for testing on application layer control in smart home environments, and network layer control including network functionalities like rate control and routing, for SDN based environments.

Multiparty computation refers to a scenario in which multiple distinct yet connected parties aim to jointly compute a functionality. Over recent decades, with the rapid spread of the internet and digital technologies, multiparty computation has become an increasingly important topic. In addition to the integrity of computation in such scenarios, it is essential to ensure that the privacy of sensitive information is not violated. Thus, secure multiparty computation aims to provide sound approaches for the joint computation of desired functionalities in a secure manner: Not only must the integrity of computation be guaranteed, but also each party must not learn anything about the other parties' private data. In other words, each party learns no more than what can be inferred from its own input and its prescribed output. This thesis considers secure two-party computation over arithmetic circuits based on additive secret sharing. In particular, we focus on efficient and secure solutions for fundamental functionalities that depend on the equality of private comparands. The first direction we take is providing efficient protocols for two major problems of interest. Specifically, we give novel and efficient solutions for private equality testing and multiple variants of secure wildcard pattern matching over any arbitrary finite alphabet. These problems are of vital importance: Private equality testing is a basic building block in many secure multiparty protocols; and, secure pattern matching is frequently used in various data-sensitive domains, including (but not limited to) private information retrieval and healthcare-related data analysis. The second direction we take towards a performance improvement in equality-based secure two-party computation is via introducing a generic functionality-independent secure preprocessing that results in an overall computation and communication cost reduction for any subsequent protocol. We achieve this by providing the first precise functionality formulation and secure protocols for replacing original inputs with much smaller inputs such that this replacement neither changes the outcome of subsequent computations nor violates the privacy of sensitive inputs. Moreover, our input-size reduction opens the door to a new approach for efficiently solving Private Set Intersection. The protocols we give in this thesis are typically secure in the semi-honest adversarial threat model.

The Unmanned Aerial Vehicles (UAV) technology has evolved exponentially in recent years. Smaller and less expensive devices allow a world of new applications in different areas, but as this progress can be beneficial, the use of UAVs with malicious intentions also poses a threat. UAVs can carry weapons or explosives and access restricted zones passing undetected, representing a real threat for civilians and institutions. Acoustic detection in combination with machine learning models emerges as a viable solution since, despite its limitations related with environmental noise, it has provided promising results on classifying UAV sounds, it is adaptable to multiple environments, and especially, it can be a cost-effective solution, something much needed in the counter UAV market with high projections for the coming years. The problem addressed by this project is the need for a real-world adaptable solution which can show that an array of acoustic sensors can be implemented for the detection and localization of UAVs with minimal cost and competitive performance. In this research, a low-cost acoustic detection system that can detect, in real time, about the presence and direction of arrival of a UAV approaching a target was engineered and validated. The model developed includes an array of acoustic sensors remotely connected to a central server, which uses the sound signals to estimate the direction of arrival of the UAV. This model works with a single microphone per node which calculates the position based on the acoustic intensity change produced by the UAV, reducing the implementation costs and being able to work asynchronously. The development of the project included collecting data from UAVs flying both indoors and outdoors, and a performance analysis under realistic conditions. The results demonstrated that the solution provides real time UAV detection and localization information to protect a target from an attacking UAV, and that it can be applied in real world scenarios.

This paper aims to present a novel approach to real-time indoor temperature forecasting to meet energy consumption constraints in buildings, utilizing computing resources available at the edge of a network, close to data sources. This work was inspired by the irreversible effects of global warming accelerated by greenhouse gas emissions from burning fossil fuels. As much as human activities have heavy impacts on global energy use, it is of utmost importance to reduce the amount of energy consumed in every possible scenario where humans are involved. According to the US Environmental Protection Agency (EPA), one of the biggest greenhouse gas sources is commercial and residential buildings, which took up 13 percent of 2019 greenhouse gas emissions in the United States. In this context, it is assumed that information of the building environment such as indoor temperature and indoor humidity, and predictions based on the information can contribute to more accurate and efficient regulation of indoor heating and cooling systems. When it comes to indoor temperature, distributed IoT devices in buildings can enable more accurate temperature forecasting and eventually help to build administrators in regulating the temperature in an energy-efficient way, but without damaging the indoor environment quality. While the IoT technology shows potential as a complement to HVAC control systems, the majority of existing IoT systems integrate a remote cloud to transfer and process all data from IoT sensors. Instead, the proposed IoT system incorporates the concept of edge computing by utilizing small computer power in close proximity to sensors where the data are generated, to overcome problems of the traditional cloud-centric IoT architecture. In addition, as the microcontroller at the edge supports computing, the machine learning-based prediction of indoor temperature is performed on the microcomputer and transferred to the cloud for further processing. The machine learning algorithm used for prediction, ANN (Artificial Neural Network) is evaluated based on error metrics and compared with simple prediction models.

Anonymous communication networks (ACNs) are critical to communication privacy over the internet as they enableindividuals to maintain their privacy from untrusted intermediaries and endpoints.Typically, ACNs involve messages traveling through some intermediaries before arriving at their destinations, and therefore they introduce network latency and bandwidth overheads. The goal of this work is to investigate the fundamental constraints of anonymous communication (AC) protocols.We analyze the relationship between bandwidth overhead, latency overhead, and sender anonymity or recipient anonymity against a global passive (network-level) adversary. We confirm the widely believed trilemma that an AC protocol can only achieve two out of the following three properties: strong anonymity (i.e., anonymity up to a negligible chance),low bandwidth overhead, and low latency overhead. We further study anonymity against a stronger global passive adversary that can additionally passively compromise some of the AC protocol nodes.For a given number of compromised nodes, we derive as a necessary constraint a relationship between bandwidth and latency overhead whose violation make it impossible for an AC protocol to achieve strong anonymity. We analyze prominent AC protocols from the literature and depict to which extent those satisfy our necessary constraints. Our fundamental necessary constraints offer a guideline not only for improving existing AC systems but also for designing novel AC protocols with non-traditional bandwidth and latency overhead choices.Using the guidelines indicated by our fundamental necessary constraints we provide two efficient protocol constructions.First, we design a mixnet-based AC protocol Streams that provides provable mixing guarantees with the expense of latency overhead. Streams realizes a trusted third party stop-and-go mix as long as each message stays in the system for $\omega(\log \eta)$ rounds.Second, we offer a DC-net based design OrgAn that can provide strong sender anonymity with constant latency at the expense of bandwidth overhead. OrgAn solves the problem of regular requirements of key and slot agreement present in typical DC-net based protocols, by utilizing a client/relay/server architecture.

An RGBZ synthetic dataset consisting of five object classes in a variety of virtual environments and orientations was combined with a small sample of real-world image data and used to train the Mask R-CNN (MR-CNN) architecture in a variety of configurations. When the MR-CNN architecture was initialized with MS COCO weights and the heads were trained with a mix of synthetic data and real world data, F1 scores improved in four of the five classes: The average maximum F1-score of all classes and all epochs for the networks trained with synthetic data is F1∗ = 0.91, compared to F1 = 0.89 for the networks trained exclusively with real data, and the standard deviation of the maximum mean F1-score for synthetically trained networks is σ∗ F1 = 0.015, compared to σF 1 = 0.020 for the networks trained exclusively with real data. Various backgrounds in synthetic data were shown to have negligible impact on F1 scores, opening the door to abstract backgrounds and minimizing the need for intensive synthetic data fabrication. When the MR-CNN architecture was initialized with MS COCO weights and depth data was included in the training data, the net- work was shown to rely heavily on the initial convolutional input to feed features into the network, the image depth channel was shown to influence mask generation, and the image color channels were shown to influence object classification. A set of latent variables for a subset of the synthetic datatset was generated with a Variational Autoencoder then analyzed using Principle Component Analysis and Uniform Manifold Projection and Approximation (UMAP). The UMAP analysis showed no meaningful distinction between real-world and synthetic data, and a small bias towards clustering based on image background.

The generalization performance of neural networks can be improved by respecting invariances inherent in a problem. For instance, it is often the case that rotating an imagedoes not change its meaning; the target variable is likely invariant to rotations of the input. Models for graph and set data are typically designed to be permutation-invariant, since the order of set elements and isomorphisms of graphs usually do not change their meaning, but a unified paradigm for both types of data is lacking. Moreover, existing models are either insufficiently flexible or difficult to reliably train in practice. This dissertation presents Janossy pooling (JP), a novel framework for training neural networks that are (approximately) invariant to a prespecified finite group of transformations (e.g., permutations), with a focus on graphs and sets. Motivated by partial exchangeability and Janossy densities, we view transformation-invariant models as averages over all transformations of the input, each passed to a transformation-sensitive function. As the set of transformations can be very large, we advance three approximation strategies: π-Stochastic Gradient Descent (π-SGD), k-ary approximations, and poly-canonical orderings. Compared to state-of-the-art approaches, JP is capable of expressing a richer class of models than Message Passing Graph Neural Networks and does not necessitate the use of highly complex and discontinuous functions for modeling sets effectively. Empirical evidence, including experiments on molecular and protein-protein-interaction datasets, supports the theory we develop and the practical benefits of JP. However, on another note, it may be unclear whether enforcing invariances is appropriate in any given task, such as in the case of time-evolving graphs.Accordingly, we propose a data-driven scheme where the extent of transformation sensitivity is determined by a regularization hyperparameter. This approach invokes the Birkhoff-von Neumann theorem, making it directly applicable to both graphs and sets, and establishes a link between existing methods and infinitely strong regularization. Finally, we verify this approach experimentally.

With more and more videos taken from dash cams on thousands of cars, retrieving these videos and searching for important information is a daunting task. The purpose of this work is to mine some key road and vehicle motion attributes in a large-scale driving video data set for traffic analysis, sensing algorithm development and autonomous driving test benchmarks. Current sensing and control of autonomous cars based on full-view identification makes it difficult to maintain a high-frequency with a fast-moving vehicle, since computation is increasingly used to cope with driving environment changes.A big challenge in video data mining is how to deal with huge amounts of data. We use a compact representation called the road profile system to visualize the road environment in long 2D images. It reduces the data from each frame of image to one line, thereby compressing the video clip to the image. This data dimensionality reduction method has several advantages: First, the data size is greatly compressed. The data is compressed from a video to an image, and each frame in the video is compressed into a line. The data size is compressed hundreds of times. While the size and dimensionality of the data has been compressed greatly, the useful information in the driving video is still completely preserved, and motion information is even better represented more intuitively. Because of the data and dimensionality reduction, the identification algorithm computational efficiency is higher than the full-view identification method, and it makes the real-time identification on road is possible. Second, the data is easier to be visualized, because the data is reduced in dimensionality, and the three-dimensional video data is compressed into two-dimensional data, the reduction is more conducive to the visualization and mutual comparison of the data. Third, continuously changing attributes are easier to show and be captured. Due to the more convenient visualization of two-dimensional data, the position, color and size of the same object within a few frames will be easier to compare and capture. At the same time, in many cases, the trouble caused by tracking and matching can be eliminated. Based on the road profile system, there are three tasks in autonomous driving are achieved using the road profile images.The first application is road edge detection under different weather and appearance for road following in autonomous driving to capture the road profile image and linearity profile image in the road profile system. This work uses naturalistic driving video data mining to study the appearance of roads, which covers large-scale road data and changes. This work excavated a large number of naturalistic driving video sets to sample the light-sensitive area for color feature distribution. The effective road contour image is extracted from the long-time driving video, thereby greatly reducing the amount of video data. Then, the weather and lighting type can be identified. For each weather and lighting condition obvious features are I identified at the edge of the road to distinguish the road edge. The second application is detecting vehicle interactions in driving videos via motion profile images to capture the motion profile image in the road profile system. This work uses visual actions recorded in driving videos taken by a dashboard camera to identify this interaction. The motion profile images of the video are filtered at key locations, thereby reducing the complexity of object detection, depth sensing, target tracking and motion estimation. The purpose of this reduction is for decision making of vehicle actions such as lane changing, vehicle following, and cut-in handling.The third application is motion planning based on vehicle interactions and driving video. Taking note of the fact that a car travels in a straight line, we simply identify a few sample lines in the view to constantly scan the road, vehicles, and environment, generating a portion of the entire video data. Without using redundant data processing, we performed semantic segmentation to streaming road profile images. We plan the vehicle's path/motion using the smallest data set possible that contains all necessary information for driving.The results are obtained efficiently, and the accuracy is acceptable. The results can be used for driving video mining, traffic analysis, driver behavior understanding, etc.

The progress of computer vision has extended the visual recognition tasks from images to the video domain. In recent years, multi-object tracking and segmentation (MOTS), which instantly performs track and segment multiple objects in video frames, has arrested much research attention as it holds valuable potential for emerging technology like autonomous driving. In this work, I proposed an effective framework for the MOTS task on live-streaming video frames, for simultaneously detecting objects, segmenting instances, and tracking move- ments across frames in an end-to-end learnable fashion. The core idea of the proposed flow- guided association (FGA) method was to leverage flow fields to increase pixel-level feature representation and used associative connections across individual task heads (a.k.a. detec- tion head, segmentation head, and tracking head) to facilitate feature sharing. The novel connection architectures allow upper-level tasks (i.e. segmentation and tracking) can accu- rately fire on the region of interests (RoI) for the final prediction. Extensive evaluations on KITTI MOTS dataset, MOTS-Challenge dataset, and new DGL-MOTS dataset indicated that the proposed method was competitive with the best existing method.In addition, the current MOTS dataset inadequately captures the real-world complexity to train a deep-learning algorithm to address the extensive varieties of driving settings. To address the deficiency, I presented the DGL-MOTS Dataset and DG-Labeler for data an- notation for MOTS work. DGL-MOTS included 106,089 instance masks for 1,632 distinct objects in 40 video recordings. My effort exceeded the state-of-the-art KITTI MOTS dataset in terms of dataset scale, object density and variations, and scene diversity. Results on ex- tensive cross-dataset evaluations indicated significant performance improvements for several state-of-the-art methods trained on our DGL-MOTS dataset. I believe the FGA algorithm and the DGL-MOTS Dataset hold valuable potential to facilitate the progress of MOTS studies for autonomous driving.The progress of computer vision has extended the visual recognition tasks from images to the video domain. In recent years, multi-object tracking and segmentation (MOTS), which instantly performs track and segment multiple objects in video frames, has arrested much research attention as it holds valuable potential for emerging technology like autonomous driving. In this work, I proposed an effective framework for the MOTS task on live-streaming video frames, for simultaneously detecting objects, segmenting instances, and tracking movements across frames in an end-to-end learnable fashion. The core idea of the proposed flow-guided association (FGA) method was to leverage flow fields to increase pixel-level feature representation and used associative connections across individual task heads (a.k.a. detection head, segmentation head, and tracking head) to facilitate feature sharing. The novel connection architectures allow upper-level tasks (i.e. segmentation and tracking) can accurately fire on the region of interests (RoI) for the final prediction. Extensive evaluations on KITTI MOTS dataset, MOTS-Challenge dataset, and new DGL-MOTS dataset indicated that the proposed method was competitive with the best existing method.

Electronic health records collect an enormous amount of data about patients. However, the information about the patient’s illness is stored in progress notes that are in an un- structured format. It is difficult for humans to annotate symptoms listed in the free text. Recently, researchers have explored the advancements of deep learning can be applied to pro- cess biomedical data. The information in the text can be extracted with the help of natural language processing. The research presented in this thesis aims at automating the process of symptom extraction. The proposed methods use pre-trained word embeddings such as BioWord2Vec, BERT, and BioBERT to generate vectors of the words based on semantics and syntactic structure of sentences. BioWord2Vec embeddings are fed into a BiLSTM neural network with a CRF layer to capture the dependencies between the co-related terms in the sentence. The pre-trained BERT and BioBERT embeddings are fed into the BERT model with a CRF layer to analyze the output tags of neighboring tokens. The research shows that with the help of the CRF layer in neural network models, longer phrases of symptoms can be extracted from the text. The proposed models are compared with the UMLS Metamap tool that uses various sources to categorize the terms in the text to different semantic types and Stanford CoreNLP, a dependency parser, that analyses syntactic relations in the sentence to extract information. The performance of the models is analyzed by using strict, relaxed, and n-gram evaluation schemes. The results show BioBERT with a CRF layer can extract the majority of the human-labeled symptoms. Furthermore, the model is used to extract symptoms from COVID-19 tweets. The model was able to extract symptoms listed by CDC as well as new symptoms.

Severe acute respiratory syndrome (SARS)-CoV-2 emerged in late 2019, then became an unprecedented public health crisis. Hundreds of millions of people have been affected. What is worse, many researchers have revealed that COVID-19 may have long-term effects on varieties of organs even after recovery. Consequently, there is a need for the study of its sequelae. The purpose of this project is to use machine learning algorithms to study the relationship between patients��� EMR data and long-term sequelae, especially kidney diseases. Inspired by a recent learning disentangled representation for recommendation work, this project proposes a method that (i) predicts the development trend of the kidney disease; (ii) learn representations that uncover and disentangle factors related to kidney diseases. The major contribution is that this model has high interpretability which enables medical works to infer the development of patients' condition.

Federated learning (FL) is a decentralized privacy-preserving learning technique in which clients learn a joint collaborative model through a central aggregator without sharing their data. In this setting, all clients learn a single common predictor (FedAvg), which does not generalize well on each client's local data due to the statistical data heterogeneity among clients. In this paper, we address this problem with PersFL, a discrete two-stage personalized learning algorithm. In the first stage, PersFL finds the optimal teacher model of each client during the FL training phase. In the second stage, PersFL distills the useful knowledge from optimal teachers into each user's local model. The teacher model provides each client with some rich, high-level representation that a client can easily adapt to its local model, which overcomes the statistical heterogeneity present at different clients. We evaluate PersFL on CIFAR-10 and MNIST datasets using three data-splitting strategies to control the diversity between clients' data distributions.We empirically show that PersFL outperforms FedAvg and three state-of-the-art personalization methods, pFedMe, Per-FedAvg and FedPer on majority data-splits with minimal communication cost. Further, we study the performance of PersFL on different distillation objectives, how this performance is affected by the equitable notion of fairness among clients, and the number of required communication rounds. We also build an evaluation framework with the following modules: Data Generator, Federated Model Generation, and Evaluation Metrics. We introduce new metrics for the domain of personalized FL, and split these metrics into two perspectives: Performance, and Fairness. We analyze the performance of all the personalized algorithms by applying these metrics to answer the following questions: Which personalization algorithm performs the best in terms of accuracy across all the users?, and Which personalization algorithm is the fairest amongst all of them? Finally, we make the code for this work available at https://tinyurl.com/1hp9ywfa for public use and validation.

Indiana University-Purdue University Indianapolis (IUPUI), Recommender systems are widely used in many domains. In this work, the importance of a recommender system in an online learning platform is discussed. After explaining the concept of adding an intelligent agent to online education systems, some features of the Course Networking (CN) website are demonstrated. Finally, the relation between CN, the intelligent agent (Rumi), and the recommender system is presented. Along with the argument of three different approaches for building a community recommendation system. The result shows that the Neighboring Collaborative Filtering (NCF) outperforms both the transfer learning method and the Continuous bag-of-words approach. The NCF algorithm has a general format with two various implementations that can be used for other recommendations, such as course, skill, major, and book recommendations.

The detection of anomalies in real-world networks is applicable in different domains; the application includes, but is not limited to, credit card fraud detection, malware identification and classification, cancer detection from diagnostic reports, abnormal traffic detection, identification of fake media posts, and the like. Many ongoing and current researches are providing tools for analyzing labeled and unlabeled data; however, the challenges of finding anomalies and patterns in large-scale datasets still exist because of rapid changes in the threat landscape. In this study, I implemented a novel and robust solution that combines data science and cybersecurity to solve complex network security problems. I used Long Short-Term Memory (LSTM) model, Louvain algorithm, and PageRank algorithm to identify and group anomalies in large-scale real-world networks. The network has billions of packets. The developed model used different visualization techniques to provide further insight into how the anomalies in the network are related. Mean absolute error (MAE) and root mean square error (RMSE) was used to validate the anomaly detection models, the results obtained for both are 5.1813e-04 and 1e-03 respectively. The low loss from the training phase confirmed the low RMSE at loss: 5.1812e-04, mean absolute error: 5.1813e-04, validation loss: 3.9858e-04, validation mean absolute error: 3.9858e-04. The result from the community detection shows an overall modularity value of 0.914 which is proof of the existence of very strong communities among the anomalies. The largest sub-community of the anomalies connects 10.42% of the total nodes of the anomalies. The broader aim and impact of this study was to provide sophisticated, AI-assisted countermeasures to cyber-threats in large-scale networks. To close the existing gaps created by the shortage of skilled and experienced cybersecurity specialists and analysts in the cybersecurity field, solutions based on out-of-the-box thinking are inevitable; this research was aimed at yielding one of such solutions. It was built to detect specific and collaborating threat actors in large networks and to help speed up how the activities of anomalies in any given large-scale network can be curtailed in time.

Today, machine learning models and Deep Neural Networks (DNNs) are prevalent in various areas. Also, educational Artificial Intelligence (AI) is drawing increasing attentionwith the rapid development of online learning platforms. Researchers explore different types of educational AI to improve students��� learning performance and experience in online classes. Educational AIs can be categorized into ���interactive��� and ���predictive.��� Interactive AIs answer simple course questions for students, such as the due day of homework and the final project���s minimum page requirement. Predictive educational AIs play a role in predicting students��� learning states. Instructors can adjust the learning content based on the students��� learning states. However, most AIs are not evaluated in an actual class setting. Therefore, we want to evaluate the effects of a state-of-the-art educational AI model, BKT (Bayesian Knowledge Tracing)-LSTM(Long Short-Term Memory), on students��� learning performance in an actual class setting. Data came from the course CNIT 25501, a large introductory Java program?ming class at Purdue University. Participants were randomly separated into the control and experimental groups (AI-group). Weekly quizzes measured participants��� learning performance. Pre-quiz and base quizzes estimated participants��� prior knowledge levels. Using BKT-LSTM, participants in the experimental group had questions from the knowledge that they were most lacking. However, participants in the control group had questions from randomly picked knowledge. The results suggested that both the experimental and control groups had lower scores in review quizzes than in base quizzes. However, the score difference between base quizzes and review quizzes for the experimental group was more often significantly different (three quizzes) compared to the control group (two quizzes), demonstrating the predictive capability of BKT-LSTM to some extent. Initially, we expected that BKT-LSTM would enhance students��� learning performance. However, in post-quiz, participants in the control group had significantly higher scores than those in the experimental group. The result suggested that continuous complex questions may negatively affect students��� learning initiatives. On the contrary, relatively easy questions may improve their learning initiatives.

With the advance of the technologies of Internet of things, smart devices or virtual personal assistants at home, such as Google Assistant, Apple Siri, and Amazon Alexa, have been widely used to control and access different objects like door lock, blobs, air conditioner, and even bank accounts, which makes our life convenient. Because of its ease for operations, voice control becomes a main interface between users and these smart devices. To make voice control more secure, speaker verification systems have been researched to apply human voice as biometrics to accurately identify a legitimate user and avoid the illegal access. In recent studies, however, it has been shown that speaker verification systems are vulnerable to different security attacks such as replay, voice cloning, and adversarial attacks. Among all attacks, adversarial attacks are the most dangerous and very challenging to defend. Currently, there is no known method that can effectively defend against such an attack in speaker verification systems. The goal of this project is to design and implement a defense system that is simple, light-weight, and effectively against adversarial attacks for speaker verification. To achieve this goal, we study the audio samples from adversarial attacks in both the time domain and the Mel spectrogram, and find that the generated adversarial audio is simply a clean illegal audio with small perturbations that are similar to white noises, but well-designed to fool speaker verification. Our intuition is that if these perturbations can be removed or modified, adversarial attacks can potentially loss the attacking ability. Therefore, we propose to add a plugin-function module to preprocess the input audio before it is fed into the verification system. As a first attempt, we study two opposite plugin functions: denoising that attempts to remove or reduce perturbations and noise-adding that adds small Gaussian noises to an input audio. We show through experiments that both methods can significantly degrade the performance of a state-of-the-art adversarial attack. Specifically, it is shown that denoising and noise-adding can reduce the targeted attack success rate of the attack from 100% to only 56% and 5.2%, respectively. Moreover, noise-adding can slow down the attack 25 times in speed and has a minor effect on the normal operations of a speaker verification system. Therefore, we believe that noise-adding can be applied to any speaker verification system against adversarial attacks. To the best of our knowledge, this is the first attempt in applying the noise-adding method to defend against adversarial attacks in speaker verification systems.

Defending models against Natural Language Processing adversarial attacks is a challenge because of the discrete nature of the text dataset. However, given the variety of Natural Language Processing applications, it is important to make text processing models more robust and secure. This paper aims to develop techniques that will help text processing models such as BERT to combat adversarial samples that contain misspellings. These developed models are more robust than off the shelf spelling checkers.

A degree-constrained subgraph construction (DCS) problem aims to find an optimal spanning subgraph (w.r.t an objective function) subject to certain degree constraints on the vertices. DCS generalizes many combinatorial optimization problems such as Matchings and Edge Covers and has many practical and real-world applications. This thesis focuses on DCS problems where there are only upper and lower bounds on the degrees, known as b-matching and b-edge cover problems, respectively. We explore linear and submodular functions as the objective functions of the subgraph construction.The contributions of this thesis involve both the design of new approximation algorithms for these DCS problems, and also their applications to real-world contexts.We designed, developed, and implemented several approximation algorithms for DCS problems. Although some of these problems can be solved exactly in polynomial time, often these algorithms are expensive, tedious to implement, and have little to no concurrency. On the contrary, many of the approximation algorithms developed here run in nearly linear time, are simple to implement, and are concurrent. Using the local dominance framework, we developed the first parallel algorithm submodular b-matching. For weighted b-edge cover, we improved the classic Greedy algorithm using the lazy evaluation technique. We also propose and analyze several approximation algorithms using the primal-dual linear programming framework and reductions to matching. We evaluate the practical performance of these algorithms through extensive experimental results.The second contribution of the thesis is to utilize the novel algorithms in real-world applications. We employ submodular b-matching to generate a balanced task assignment for processors to build Fock matrices in the NWChemEx quantum chemistry software. Our load-balanced assignment results in a four-fold speedup per iteration of the Fock matrix computation and scales to 14,000 cores of the Summit supercomputer at Oak Ridge National Laboratory. Using approximate b-edge cover, we propose the first shared-memory and distributed-memory parallel algorithms for the adaptive anonymity problem. Minimum weighted b-edge cover and maximum weight b-matching are shown to be applicable to constructing graphs from datasets for machine learning tasks. We provide a mathematical optimization framework connecting the graph construction problem to the DCS problem.

Indiana University-Purdue University Indianapolis (IUPUI), The number of smart mobile devices has grown at a significant rate in recent years. This growth has resulted in an exponential number of publicly available mobile Apps. To help the selection of suitable Apps, from various offered choices, the App distribution platforms generally rank/recommend Apps based on average star ratings, the number of installs, and associated reviews ― all the external factors of an App. However, these ranking schemes typically tend to ignore critical internal factors (e.g., bugs, security vulnerabilities, and data leaks) of the Apps. The AppStores need to incorporate a holistic methodology that includes internal and external factors to assign a level of trust to Apps. The inclusion of the internal factors will describe associated potential security risks. This issue is even more crucial with newly available Apps, for which either user reviews are sparse, or the number of installs is still insignificant. In such a scenario, users may fail to estimate the potential risks associated with installing Apps that exist in an AppStore. This dissertation proposes a security-related and evidence-based ranking framework, called SERS (Security-related and Evidence-based Ranking Scheme) to compare similar Apps. The trust associated with an App is calculated using both internal and external factors (i.e., security flaws and user reviews) following an evidence-based approach and applying subjective logic principles. The SERS is formalized and further enhanced in the second part of this dissertation, resulting in its enhanced version, called as E-SERS (Enhanced SERS). These enhancements include an ability to integrate any number of sources that can generate evidence for an App and consider the temporal aspect and reputation of evidence sources. Both SERS and E-SERS are evaluated using publicly accessible Apps from the Google PlayStore and the rankings generated by them are compared with prevalent ranking techniques such as the average star ratings and the Google PlayStore Rankings. The experimental results indicate that E-SERS provides a comprehensive and holistic view of an App when compared with prevalent alternatives. E-SERS is also successful in identifying malicious Apps where other ranking schemes failed to address such vulnerabilities. In the third part of this dissertation, the E-SERS framework is used to propose a trust-aware composition model at two different granularities. This model uses the trust score computed by E-SERS, along with the probability of an App belonging to the malicious category, as the desired attributes for selecting a composition as the two granularities. Finally, the trust-aware composition model is evaluated with the average star rating parameter and the trust score. A holistic approach, as proposed by E-SERS, to computer a trust score will benefit all kinds of Apps including newly published Apps that follow proper security measures but initially struggle in the AppStore rankings due to a lack of a large number of reviews and ratings. Hence, E-SERS will be helpful both to the developers and users. In addition, the composition model that uses such a holistic trust score will enable system integrators to create trust-aware distributed systems for their specific needs.

Container adoption has exploded in recent years with over 92% of companies using containers as part of their cloud infrastructure. This explosion is partly due to the easy orchestration and lightweight operations of containers compared to traditional virtual machines. As container adoption increases, servers hosting containers become more attractive targets for adversaries looking to gain control of a container to steal trade secrets, exfiltrate customer data, or hijack hardware for cryptocurrency mining. To control a container host, an adversary can exploit a vulnerability that enables them to escape from the container onto the host. This kind of attack is termed a “container escape” because the adversary is able to execute code on the host from within the isolated container. The vulnerabilities which allow container escape exploits originate from three main sources: (1) container profile misconfiguration, (2) the host’s Linux kernel, and (3) the container runtime. While the first two cases have been studied in the literature, to the best of the author’s knowledge, there is, at present, no work that investigates the impact of container runtime vulnerabilities. To fill this gap, a survey over container runtime vulnerabilities was conducted investigating 59 CVEs for 11 different container runtimes. As CVE data alone would limit the investigation analysis, the investigation focused on the 28 CVEs with publicly available proof of concept (PoC) exploits. To facilitate this analysis, each exploit was broken down into a series of high-level commands executed by the adversary called “steps”. Using the steps of each CVE’s corresponding exploit, a seven-class taxonomy of these 28 vulnerabilities was constructed revealing that 46% of the CVEs had a PoC exploit which enabled a container escape. Since container escapes were the most frequently occurring category, the nine corresponding PoC exploits were further analyzed to reveal that the underlying cause of these container escapes was a host component leaking into the container. This survey provides new insight into system vulnerabilities exposed by container runtimes thereby informing the direction of future research.

Disease subtyping has been a critical aim of precision and personalized medicine. With the potential to improve patient outcomes, unsupervised and semi-supervised methods for determining phenotypes of subtypes have emerged with a recent focus on matrix and tensor factorization. However, interpretability of proposed models is debatable. Principal component analysis (PCA), a traditional method of dimensionality reduction, does not impose non-negativity constraints. Thus coefficients of the principal components are, in cases, difficult to translate to real physical units. Non-negative matrix factorization (NMF) constrains the factorization to positive numbers such that representative types resulting from the factorization are additive. Archetypal analysis (AA) extends this idea and seeks to identify pure types, archetypes, at the extremes of the data from which all other data can be expressed as a convex combination, or by proportion, of the archetypes. Using AA, this study sought to evaluate the sufficiency of AKI staging criteria through unsupervised subtyping. Archetype analysis failed to find a direct 1:1 mapping of archetypes to physician staging and also did not provide additional insight into patient outcomes. Several factors of the analysis such as quality of the data source and the difficulty in selecting features contributed to the outcome. Additionally, after performing feature selection with lasso across data subsets, it was determined that current staging criteria is sufficient to determine patient phenotype with serum creatinine at time of diagnosis to be a necessary factor.

Authoring and Controlling virtual crowds is an important function in many domains such as Computer Graphics, Civil Engineering, and robotics. Creating a virtual crowd can be achieved by a combination of authoring tools and methods, crowd simulation algorithms, and rendering animation. However, to achieve desired virtual crowd motion and agent-based optimization the current methods and tools have limitations, that prevent low-skilled and medium-skilled users from taking advantage of them.This dissertation formulates a novel Interactive Crowd Content Creation (ICCC) pipeline with a set of novel components and enhancements on existing components; User Observation and Sketching, Retargeting, Crowd Simulation, Animation, and Agent-based Optimization. The Sketching component of this pipeline introduces a novel sketching language validated by an extensive user study and analysis. The retargeting components allow to mix and match virtual and real trajectories, and interactively modify crowd motion spatially and temporally. An exploration-based scheme is introduced on top of the crowd simulation component to introduce the ability to model a crowd that does not have prior knowledge of their surroundings. The enhancements on the agent-based optimization allow for an accelerated scheme that allows running many simulations in parallel, model indicators, and run forward and inverse optimizations. Two applications are presented in this regard; One covering the domain of Urban Walkability and another covering the domain of Swarm Robotics. Altogether, these components and their sub-components form a comprehensive framework for authoring and controlling virtual crowds. Our results show that this novel pipeline improves the process of crowd content creation by enabling untrained users to create complex crowd motions such as the Shibuya crossing, mix and match spatial and temporal styles. We have also improved the ability to solve for better walkable outdoor areas, and navigate swarm robots in indoor areas.

Internet of Things (IoT) systems running on Microcontrollers (MCUS) have become a prominent target of remote attacks. Although deployed in security and safety critical domains, such systems lack basic mitigations against control-flow hijacking attacks. Attacks against IoT systems already enabled malicious takeover of smartphones, vehicles, unmanned aerial vehicles, and industrial control systems. The thesis introduces a systemic analysis of previous defense mitigations to secure IoT systems. Building off this systematization, we identify two main issues in IoT systems security. First, efforts to protect IoT systems are hindered by the lack of realistic benchmarks and evaluation frameworks. Second, existing solutions to protect from control-flow hijacking on the return edge are either impractical or have limited security guarantees. This thesis addresses these issues using two approaches. First, we present BenchIoT, a benchmark suite of five realistic IoT applications and an evaluation framework that enables automated and extensible evaluation of 14 metrics covering security, performance, memory usage, and energy. BenchIoT enables evaluating and comparing security mechanisms. Using BenchIoT, we show that even if two security mechanisms have similarly modest runtime overhead, one can have undesired consequences on security such as a large portion of privileged user execution. Second, we introduce Return Address Integrity (RAI), a novel security mechanism to prevent all control-flow hijacking attacks targeting return edges, without requiring special hardware. We design and implement μRAI to enforce the RAI property. Our results show μRAI has a low runtime overhead of 0.1% on average, and therefore is apractical solution for IoT systems. This thesis enables measuring the security IoT systems through standardized benchmarks and metrics. Using static analysis and runtime monitors, it prevents control-flow hijacking attacks on return edges with low runtime overhead. Combined, this thesis advances the state-of-the-art of protecting IoT systems and benchmarking its security.

Large datasets related to Indiana’s Education and Workforce development are used by various demographics such as stakeholders and decision makers in education and government, parents, teachers and employees of various companies to find trends and patterns in the data to better guide decision-making through statistical analysis. However, most of this data is scattered, textual and available in the form of excel sheets which makes it harder to look at the data from different perspectives, drill down and roll up the data and find trends and patterns in the data. Such data representation does not take into account the inherent characteristics of the user which can affect how well the user understands, perceives and interprets the data.Information dashboards used to view and navigate between visualizations of different datasets, provide a coherent, central access to all data, and make it easy to view different aspects of the system. The purpose of this research is to create a new data visualization dashboard for visualizing education and workforce data and find which design principles are applicable while designing such a dashboard for the target demographic in the education and workforce domain. This study also aims at assessing how the introduction of such a data dashboard affects the work processes and decision making of stakeholders involved in education and workforce development in the state of Indiana.User studies consisting of usability testing and semi-structured interviews with the stakeholders in education and workforce development in the state of Indiana is conducted to test the effectiveness of the dashboard. Finally, this research proposes how a regional map-based dashboard can be used as an effective method to design a data dashboard for education and workforce data for other states and other domains as well.

The growth in online goods delivery is causing a dramatic surge in urban vehicle traffic from last-mile deliveries. On the other hand, ride-sharing has been on the rise with the success of ride-sharing platforms and increased research on using autonomous vehicle technologies for routing and matching. The future of urban mobility for passengers and goods relies on leveraging new methods that minimize operational costs and environmental footprints of transportation systems. This paper considers combining passenger transportation with goods delivery to improve vehicle-based transportation. Even though the problem has been studied with model-based approaches where the dynamic model of the transportation system environment is defined, model-free approaches where the dynamics of the environment are learned by interaction have been demonstrated to be adaptable to new or erratic environment dynamics. FlexPool is a distributed model-free deep reinforcement learning algorithm that jointly serves passengers \& goods workloads by learning optimal dispatch policies from its interaction with the environment. The model-free algorithm (as opposed to a model-based one) is an algorithm which does not use the transition probability distribution (and the reward function) associated with the Markov decision process (MDP). The proposed algorithm pools passengers for a ride-sharing service and delivers goods using a multi-hop routing method. These flexibilities decrease the fleet's operational cost and environmental footprint while maintaining service levels for passengers and goods. The dispatching algorithm based on deep reinforcement learning is integrated with an efficient matching algorithm for passengers and goods. Through simulations on a realistic urban mobility platform, we demonstrate that FlexPool outperforms other model-free settings in serving the demands from passengers \& goods. FlexPool achieves 30\% higher fleet utilization and 35\% higher fuel efficiency in comparison to (i) model-free approaches where vehicles transport a combination of passengers \& goods without the use of multi-hop transit, and (ii) model-free approaches where vehicles exclusively transport either passengers or goods.

With the advancement in Unmanned Aerial Vehicles (UAV) technology, UAVs have become accessible to the public. However, recent world events have highlighted that the rapid increase of UAVs is bringing with it a threat to public privacy and security. Thus, it is important to think about how to prevent the threats of UAVs to protect our privacy and safety. This study aims to provide an alternative way to substitute an expensive system by using 2D optical sensors that can be easily utilized by people. One of the main challenges for aerial object recognition with computer vision is discriminating other flying objects from the targets, in the far distance. There are limitation to classify the flying object when it appears as a set of small black pixels on the frame. The movement feature can help the system to extract the discriminative feature, so that the classifier can classify the UAV and other objects, such as a bird. Thus, this study proposes a drone detection system using two elements of information, which are appearance information and motion information to overcome the limitation of a vision based system.

Over the years, the performance demand for graphics applications has been steadily increasing. While upgrading the hardware is one direct solution, the emergence of the new low-level and low-overhead graphics APIs like Vulkan also exposed the possibility of improving rendering performance from the bottom of software implementation.Most of the recent years’ middle- to high-end personal computers are equipped with both integrated and discrete GPUs. However, with previous graphics APIs, it is hard to put these two heterogeneous GPUs to work concurrently in the same application without tailored driver support.This thesis provides an exploration into the utilization of such heterogeneous GPUs in real-time rendering with the help of Vulkan API. This paper first demonstrates the design and implementation details for the proposed heterogeneous GPUs working model. After that, the paper presents the test of two workload offloading strategies: offloading screen space output workload to the integrated GPU and offloading asynchronous computation workload to the integrated GPU.While this study failed to obtain performance improvement through offloading screen space output workload, it is successful in validating that offloading asynchronous computation workload from the discrete GPU to the integrated GPU can improve the overall system performance. This study proves that it is possible to make use of the integrated and discrete GPUs concurrently in the same application with the help of Vulkan. And offloading asynchronous computation workload from the discrete GPU to the integrated GPU can provide up to 3-4% performance improvement with combinations like UHD Graphics 630 + RTX 2070 Max-Q and HD Graphics 630 + GTX 1050.

There are various real-world applications that involve large number of interacting agents, for e.g., viral marketing, personalized teaching, healthcare, recommendation systems, online communication platforms. However, much of the existing work in Multi-Agent Reinforcement Learning (MARL) focuses on small number of agents. The standard approaches to train a complex model for each user in a decentralized fashion are impractical for thousands of agents. Centralized learning is also infeasible due to the curse of dimensionality and exponential increase in joint representations. There is an opportunity to utilize the interactions and correlations between agents, to develop RL approaches that can scale for large number of agents. However, user interactions are typically sparse. In this dissertation, we dene Social Reinforcement Learning as a sub class of MARL for domains with large number of agents with relatively few (sparse) relations and interactions between them.We consider the important task of fake news mitigation as an example to demonstrate the real-world applicability of our proposed Social RL approaches. First, we propose a centralized Social RL approach to estimate incentives (interventions) required to promote the spread of true news in a social network|in order to mitigate the impact of fake news. We model news diffusion as a Multivariate Hawkes Process (MHP) and make interventions that are learnt via policy optimization in a Markov Decision Process (MDP). The key insight is to estimate the response a user will get from the social network upon sharing a post, as it indicates her impact on news diffusion, and will thus help in efficient allocation of incentive. Second, we develop an efficient centralized Social RL approach to address the challenges of computational complexity (associated with large number of agents), and sparse interaction data. Our key idea is to reduce the model size by dynamically clustering users based on their payoff and contribution to the goal. Lastly, the above proposed centralized approaches can be applied when the environment is fully observable to all agents, with a common system shared between all agents. To develop solutions for scenarios where agents receive only a partial view of the environment, and agents can also have separate individual goals, we propose a Social RL approach that is more decentralized. Our key idea is to use sequential parameter sharing and ego-network extrapolation to incorporate agent correlations and improvise estimates of the partially hidden system information. We evaluate our proposed approaches on two Twitter datasets. Our centralized learning methods outperform other alternatives that do not consider estimates of user feedback when learning how to allocate incentives. Furthermore, by clustering users, we are able to achieve faster convergence along with learning more accurate estimates, compared to baselines that do not model agent correlations or only use static clusters. Additionally, our decentralized learning approach achieves performance equivalent to that of centralized learning approach and superior performance to other baselines that either consider complete system information available to an agent, or other estimates of the hidden environment state.

The growth of autonomous vehicles, ridesharing systems, and self driving technology will bring a shift in the way ride hailing platforms plan out their services. However, these advances in technology coupled with road congestion, environmental concerns, fuel usage, vehicles emissions, and the high cost of the vehicle usage have brought more attention to better utilize the use of vehicles and their capacities. In this paper, we propose a novel distributed multi-hop ride-sharing (MHRS) algorithm that uses deep reinforcement learning to learn optimal vehicle dispatch and matching decisions by interacting with the external environment. By allowing customers to transfer between vehicles, i.e., ride with one vehicle for sometime and then transfer to another one, MHRS helps in attaining 30% lower cost and 20% more efficient utilization of fleets, as compared to the ride-sharing algorithms. This flexibility of multi-hop feature gives a seamless experience to customers and ride-sharing companies, and thus improves ride-sharing services.

Cities have adopted evolving urban digitization strategies, and most of those increasingly focus on data, especially in the field of public transportation. Transportation data have intuitively spatial and temporal characteristics, for they are often described with when and where the trips occur. Since a trip is often described with many attributes, the transportation data can be presented with a tensor, a container which can house data in $N$-dimensions. Unlike a traditional data frame, which only has column variables, tensor is intuitively more straightforward to explore spatio-temporal data-sets, which makes those attributes more easily interpreted. However, it requires unique techniques to extract useful and relatively correct information in attributes highly correlated with each other. This work presents a mixed model consisting of tensor decomposition combined with seasonal vector autoregression in time to find latent patterns within historical taxi data classified by types of taxis, pick-up and drop-off times of services in NYC, so that it can help predict the place and time where taxis are demanded. We validated the proposed approach using the experiment evaluation with real NYC tax data. The proposed method shows the best prediction among alternative models without geographical inference, and captures the daily patterns of taxi demands for business and entertainment needs.

Many practical applications benefit from the reconstruction of a smooth multivariate function from discrete data for purposes such as reducing file size or improving analytic and visualization performance. Among the different reconstruction methods, tensor product B-spline has a number of advantageous properties over alternative data representation. However, the problem of constructing a best-fit B-spline approximation effectively contains many roadblocks. Within the many free parameters in the B-spline model, the choice of the knot vectors, which defines the separation of each piecewise polynomial patch in a B-spline construction, has a major influence on the resulting reconstruction quality. Yet existing knot placement methods are still ineffective, computationally expensive, or impose limitations on the dataset format or the B-spline order. Moving beyond the 1D cases (curves) and onto higher dimensional datasets (surfaces, volumes, hypervolumes) introduces additional computational challenges as well. Further complications also arise in the case of undersampled data points where the approximation problem can become ill-posed and existing regularization proves unsatisfactory.This dissertation is concerned with improving the efficiency and accuracy of the construction of a B-spline approximation on discrete data. Specifically, we present a novel B-splines knot placement approach for accurate reconstruction of discretely sampled data, first in 1D, then extended to higher dimensions for both structured and unstructured formats. Our knot placement methods take into account the feature or complexity of the input data by estimating its high-order derivatives such that the resulting approximation is highly accurate with a low number of control points. We demonstrate our method on various 1D to 3D structured and unstructured datasets, including synthetic, simulation, and captured data. We compare our method with state-of-the-art knot placement methods and show that our approach achieves higher accuracy while requiring fewer B-spline control points. We discuss a regression approach to the selection of the number of knots for multivariate data given a target error threshold. In the case of the reconstruction of irregularly sampled data, where the linear system often becomes ill-posed, we propose a locally varying regularization scheme to address cases for which a straightforward regularization fails to produce a satisfactory reconstruction.

Over the last few years, there has been growing interest in learning models for various Natural Language Processing tasks, such as the popular blocks world domain and fake news detection. These works typically view these problems as single-step processes, in which a human operator gives an instruction and an automated agent is evaluated on its ability to execute it. In this work, we take the first step towards increasing the bandwidth of this interaction, and suggest a protocol for including advice, high-level observations about the task, which can help constrain the agents prediction. Advice is designed to be a short natural language sentence, provided by the human operator in addition the original input for the task, that can help the agent improve its performance. We evaluate our advice-based approach on the blocks world task and fake-news detection, and show that even simple advice can help lead to significant performance improvements. To help reduce the effort involved in supplying the advice, we also explore model self-generated advice which can still improve results.

In this work, we model latent discourse indicators to classify constructive and collaborative conversations online. Such conversations are considered good as they are rich in content and have a sense of direction to resolve an issue, solve a problem or gain new insights and knowledge. These unique discourse indicators are able to characterize flow of information, sentiment and community structure within discussions. We build a deep relational model that captures these complex discourse behaviors as latent variables and make a global prediction about overall conversation based on these higher level discourse behaviors. DRaiL, a Declarative Deep Relational Learning platform built on PyTorch, is used for our task in which relevant discourse behaviors are formulated as discrete latent variables and scored using a deep model. These variables capture the nuances involved in online conversations and provide the information needed for predicting the presence or absence of collaborative and constructive characterization in the entire conversational thread. We show that the joint modeling of such competing latent behaviors results in a performance improvement over the traditional direct classification methods in which all the raw features are just combined together to predict the final decision. The Yahoo News Annotated Comments Corpus is used as a dataset containing discussions on Yahoo news forums and final labels are annotated based on our precise and restricted definitions of positively labeled conversations. We formulated our annotation guidelines based on a sample set of conversations and resolved any conflict in specific annotation by revisiting those examples again.

The accelerated urbanization of countries has led the adoption of the smart power grid with an explosion in high power usage. The emergence of Non-intrusive load monitoring (NILM), also referred to as Energy Disaggregation has followed the recent worldwide adoption of smart meters in smart grids. NILM is a convenient process to analyze composite electrical energy load and determine electrical energy consumption.A number of state-of-the-art NILM (energy disaggregation) algorithms have been proposed recently to detect various individual appliances from one aggregated signal observation. Different kinds of classification methods such as Hidden Markov Model (HMM), Support Vector Method (SVM), neural networks, fuzzy logic, Naive Bayes, k-Nearest Neighbors (kNN), and many other hybrid approaches have been used to classify the estimated power consumption of electrical appliances from extracted appliances signatures. This study proposes an end-to-end edge computing system with an NILM algorithm, which especially focuses on recognizing Electric Vehicle (EV) charging. This system consists of three main components: (1) Data acquisition and Preprocessing, (2) Extraction of EV charging load via an NILM algorithm (Load identification) on the NILMTK Framework, (3) and Result report to the cloud server platform.The monitoring of energy consumption through the proposed system is remarkably beneficial for demand response and energy efficiency. It helps to improve the understanding and prediction of power grid stress as well as enhance grid system reliability and resilience of the power grid. Furthermore, it is highly advantageous for the integration of more renewable energies that are under rapid development. As a result, countless potential NILM use-cases are expected from monitoring and identifying energy consumption in a power grid. It would enable smarter power consumption plans for residents as well as more flexible power grid management for electric utility companies, such as Duke Energy and ComEd.

Graph-structured data is pervasive. Modeling large-scale network-structured datasets require graph processing and management systems such as graph databases. Further, the analysis of graph-structured data often necessitates bulk downloads/uploads from/to the cloud or edge nodes. Unfortunately, experience has shown that malicious actors can compromise the confidentiality of highly-sensitive data stored in the cloud or shared nodes, even in an encrypted form. For particular use cases —multi-modal knowledge graphs, electronic health records, finance— network-structured datasets can be highly sensitive and require auditability, authentication, integrity protection, and privacy-preserving computation in a controlled and trusted environment, i.e., the traditional cloud computation is not suitable for these use cases. Similarly, many modern applications utilize a "shared, replicated database" approach to provide accountability and traceability. Those applications often suffer from significant privacy issues because every node in the network can access a copy of relevant contract code and data to guarantee the integrity of transactions and reach consensus, even in the presence of malicious actors.This dissertation proposes breaking from the traditional cloud computation model, and instead ship certified pre-approved trusted code closer to the data to protect graph-structured data confidentiality. Further, our technique runs in a controlled environment in a trusted data owner node and provides proof of correct code execution. This computation can be audited in the future and provides the building block to automate a variety of real use cases that require preserving data ownership. This project utilizes trusted execution environments (TEEs) but does not rely solely on TEE's architecture to provide privacy for data and code. We thoughtfully examine the drawbacks of using trusted execution environments in cloud environments. Similarly, we analyze the privacy challenges exposed by the use of blockchain technologies to provide accountability and traceability.First, we propose AGAPECert, an Auditable, Generalized, Automated, Privacy-Enabling, Certification framework capable of performing auditable computation on private graph-structured data and reporting real-time aggregate certification status without disclosing underlying private graph-structured data. AGAPECert utilizes a novel mix of trusted execution environments, blockchain technologies, and a real-time graph-based API standard to provide automated, oblivious, and auditable certification. This dissertation includes the invention of two core concepts that provide accountability, data provenance, and automation for the certification process: Oblivious Smart Contracts and Private Automated Certifications. Second, we contribute an auditable and integrity-preserving graph processing model called AuditGraph.io. AuditGraph.io utilizes a unique block-based layout and a multi-modal knowledge graph, potentially improving access locality, encryption, and integrity of highly-sensitive graph-structured data. Third, we contribute a unique data store and compute engine that facilitates the analysis and presentation of graph-structured data, i.e., TruenoDB. TruenoDB offers better throughput than the state-of-the-art. Finally, this dissertation proposes integrity-preserving streaming frameworks at the edge of the network with a personalized graph-based object lookup.

The fact that every individual has a different sense of humor and it varies greatly from one person to another means that it is a challenge to learn any individual’s humor preferences. Humor is much more than just a source of entertainment; it is an essential tool that aids communication. Understanding humor preferences can lead to improved social interactions and bridge existing social or economic gaps. In this study, we propose a methodology that aims to develop a recommendation system for jokes by analyzing its text. Various researchers have proposed different theories of humor depending on their area of focus. This exploratory study focuses mainly on Attardo and Raskin’s (1991) General Theory of Verbal Humor and implements the knowledge resources defined by it to annotate the jokes. These annotations contain the characteristics of the jokes and also play an important role in determining how alike these jokes are. We use Lin’s similarity metric (Lin, 1998) to computationally capture this similarity. The jokes are clustered in a hierarchical fashion based on their similarity values used for the recommendation. We also compare our joke recommendations to those obtained by the Eigenstate algorithm (Goldberg, Roeder, Gupta, & Perkins, 2001), an existing joke recommendation system that does not consider the content of the joke in its recommendation.

In recent years, developers have used the proliferation of biometric sensors in smart devices, along with recent advances in deep learning, to implement an array of biometrics-based recognition systems. Though these systems demonstrate remarkable performance and have seen wide acceptance, they present unique and pressing security and privacy concerns. One proposed method which addresses these concerns is the elegant, fusion-based Biometric-Capsule (BC) scheme. The BC scheme is provably secure, privacy-preserving, cancellable and interoperable in its secure feature fusion design. In this work, we demonstrate that the BC scheme is uniquely fit to secure state-of-the-art facial verification, authentication and identification systems. We compare the performance of unsecured, underlying biometrics systems to the performance of the BC-embedded systems in order to directly demonstrate the minimal effects of the privacy-preserving BC scheme on underlying system performance. Notably, we demonstrate that, when seamlessly embedded into a state-of-the-art FaceNet and ArcFace verification systems which achieve accuracies of 97.18% and 99.75% on the benchmark LFW dataset, the BC-embedded systems are able to achieve accuracies of 95.13% and 99.13% respectively. Furthermore, we also demonstrate that the BC scheme outperforms or performs as well as several other proposed secure biometric methods.

Word embeddings are utilized in various natural language processing tasks. Although effective in helping computers learn linguistic patterns employed in natural language, word embeddings also tend to learn unwanted word associations. This affects the performance of NLP tasks, as unwanted word associations propagate and amplify biases. Current word association evaluation methods for word embeddings do not account for changes in word embedding models and training corpora, when creating the rubric for word association evaluation. Current literature also lacks a consistent training and evaluation protocol for comparison of word associations across varying word embedding models and varying training corpora. In order to address this gap in prior literature, this research aims to evaluate different types of word associations, not limited to gender, racial or religious attributes, incorporating and evaluating the diachronic and variable nature of words over text data collected over a period of 200 years. This thesis introduces a framework to track changes in word associations between neutral words (proper nouns) and attributes (adjectives), across different word embedding models, over a temporal dimension, by evaluating clustering tendencies between neutral words (proper nouns) and attributive words (adjectives) over five different word embedding frameworks: Word2vec (CBOW), Word2vec (Skip-gram), GloVe, fastText (CBOW) and fastText (Skip-gram) and 20 decades of text data from 1810s to 2000s. Finally, various cluster level and corpus level measurements will be compared across aforementioned word embedding frameworks, to find how word associations evolve with changes in the embedding model and the training corpus.

This thesis describes a data provenance framework and other associated frameworks for utilizing provenance for data quality and reproducibility. We first identify the requirements for the design of a comprehensive provenance framework which can be applicable to various applications, supports a rich set of provenance metadata, and is interoperable with other provenance management systems. We then design and develop a provenance framework, called SimP, addressing such requirements. Next, we present four prominent applications and investigate how provenance data can be beneficial to such applications. The first application is the quality assessment of access control policies. Towards this, we design and implement the ProFact framework which uses provenance techniques for collecting comprehensive data about actions which were either triggered due to a network context or a user (i.e., a human or a device) action. Provenance data are used to determine whether the policies meet the quality requirements. ProFact includes two approaches for policy analysis: structure-based and classification-based. For the structure-based approach, we design tree structures to organize and assess the policy set efficiently. For the classification-based approach, we employ several classification techniques to learn the characteristics of policies and predict their quality. In addition, ProFact supports policy evolution and the assessment of its impact on the policy quality. The second application is workflow reproducibility. Towards this, we implement ProWS which is a provenance-based architecture for retrieving workflows. Specifically, ProWS transforms data provenance into workflows and then organizes data into a set of indexes to support efficient querying mechanisms. ProWS supports composite queries on three types of search criteria: keywords of workflow tasks, patterns of workflow structure, and metadata about workflows (e.g., how often a workflow was used). The third application is the access control policy reproducibility. Towards this, we propose a novel framework, Polisma, which generates attribute-based access control policies from data, namely from logs of historical access requests and their corresponding decisions. Polisma combines data mining, statistical, and machine learning techniques, and capitalizes on potential context information obtained from external sources (e.g., LDAP directories) to enhance the learning process. The fourth application is the policy reproducibility by utilizing knowledge and experience transferability. Towards this, we propose a novel framework, FLAP, which transfer attribute-based access control policies between different parties in a collaborative environment, while considering the challenges of minimal sharing of data and support policy adaptation to address conflict. All frameworks are evaluated with respect to performance and accuracy.

This paper addresses the minimum weight edge cover problem (MEC), which is stated as follows: Given a graph G= (V,E), find a set of edges S:S⊆E and ∑e∈Sw(e) ∑e∈Qw(e)∀Q: Q is an edge cover. Where an edge cover P is a set of edges such that ∀v∈V v is incident to at least one edge in P. An efficient implementation of a 4/3-approximation for MEC is provided. Empirical results obtained experimentally from practical data sets are reported and compared against various other approximation algorithms for MEC.

Real-time social media data can provide useful information on evolving events and situations. In addition, various domain users are increasingly leveraging real-time social media data to gain rapid situational awareness. Informed by discussions with first responders and government officials, we focus on two major barriers limiting the widespread adoption of social media for situational awareness: the lack of geotagged data and the deluge of irrelevant information during events. Geotags are naturally useful, as they indicate the location of origin and provide geographic context. Only a small portion of social media is geotagged, however, limiting its practical use for situational awareness. The deluge of irrelevant data provides equal difficulties, impeding the effective identification of semantically relevant information. Existing methods for short text relevance classification fail to incorporate users' knowledge into the classification process. Therefore, classifiers cannot be interactively retrained for specific events or user-dependent needs in real-time, limiting situational awareness. In this work, we first adapt, improve, and evaluate a state-of-the-art deep learning model for city-level geolocation prediction, and integrate it with a visual analytics system tailored for real-time situational awareness. We then present a novel interactive learning framework in which users rapidly identify relevant data by iteratively correcting the relevance classification of tweets in real-time. We integrate our framework with the extended Social Media Analytics and Reporting Toolkit (SMART) 2.0 system, allowing the use of our interactive learning framework within a visual analytics system adapted for real-time situational awareness.

Although current augmented, virtual, and mixed reality (AR/VR/MR) systems are facing advanced and immersive experience in the entertainment industry with countless media forms. Theses systems suffer a lack of correct direct and indirect illumination modeling where the virtual objects render with the same lighting condition as the real environment. Some systems are using baked GI, pre-recorded textures, and light probes that are mostly accomplished offline to compensate for precomputed real-time global illumination (GI). Thus, illumination information can be extracted from the physical scene for interactively rendering the virtual objects into the real world which produces a more realistic final scene in real-time. This work approaches the problem of visual coherence in AR by proposing a system that detects the real-world lighting conditions in dynamic scenes, then uses the extracted illumination information to render the objects added to the scene. The system covers several major components to achieve a more realistic augmented reality outcome. First, the detection of the incident light (direct illumination) from the physical scene with the use of computer vision techniques based on the topological structural analysis of 2D images using a live-feed 360o camera instrumented on an AR device that captures the entire radiance map. Also, the physics-based light polarization eliminates or reduces false-positive lights such as white surfaces, reflections, or glare which negatively affect the light detection process. Second, the simulation of the reflected light (indirect illumination) that bounce between the real-world surfaces to be rendered into the virtual objects and reflect their existence in the virtual world. Third, defining the shading characteristic/properties of the virtual object to depict the correct lighting assets with a suitable shadow casting. Fourth, the geometric properties of real-scene including plane detection, 3D surface reconstruction, and simple meshing are incorporated with the virtual scene for more realistic depth interactions between the real and virtual objects. These components are developed methods which assumed to be working simultaneously in real-time for photo-realistic AR. The system is tested with several lighting conditions to evaluate the accuracy of the results based on the error incurred between the real/virtual objects casting shadow and interactions. For system efficiency, the rendering time is compared with previous works and research. Further evaluation of human perception is conducted through a user study. The overall performance of the system is investigated to reduce the cost to a minimum.

Many real-world domains are relational, consisting of objects (e.g., users and pa- pers) linked to each other in various ways. Because class labels in graphs are often only available for a subset of the nodes, semi-supervised learning for graphs has been studied extensively to predict the unobserved class labels. For example, we can pre- dict political views in a partially labeled social graph dataset and get expected gross incomes of movies in an actor/movie graph with a few labels. Recently, advances in representation learning for graph data have made great strides for the semi-supervised node classification. However, most of the methods have mainly focused on learning node representations by considering simple relational properties (e.g., random walk) or aggregating nearby attributes, and it is still challenging to learn complex inter- action patterns in partially labeled graphs and provide explanations on the learned representations. In this dissertation, multiple methods are proposed to alleviate both challenges for semi-supervised node classification. First, we propose a graph neural network architecture, REGNN, that leverages local inferences for unlabeled nodes. REGNN performs graph convolution to enable label propagation via high-order paths and predicts class labels for unlabeled nodes. In particular, our proposed attention layer of REGNN measures the role equivalence among nodes and effectively reduces the noise, which is generated during the aggregation of observed labels from distant neighbors at various distances. Second, we also propose a neural network archi- tecture that jointly captures both temporal and static interaction patterns, which we call Temporal-Static-Graph-Net (TSGNet). The architecture learns a latent rep- resentation of each node in order to encode complex interaction patterns. Our key insight is that leveraging both a static neighbor encoder, that learns aggregate neigh- bor patterns, and a graph neural network-based recurrent unit, that captures complex interaction patterns, improves the performance of node classification. Lastly, in spite of better performance of representation learning on node classification tasks, neural network-based representation learning models are still less interpretable than the pre- vious relational learning models due to the lack of explanation methods. To address the problem, we show that nodes with high bridgeness scores have larger impacts on node embeddings such as DeepWalk, LINE, Struc2Vec, and PTE under perturbation. However, it is computationally heavy to get bridgeness scores, and we propose a novel gradient-based explanation method, GRAPH-wGD, to find nodes with high bridgeness efficiently. In our evaluations, our proposed architectures (REGNN and TSGNet) for semi-supervised node classification consistently improve predictive performance on real-world datasets. Our GRAPH-wGD also identifies important nodes as global explanations, which significantly change both predicted probabilities on node classification tasks and k-nearest neighbors in the embedding space after perturbing the highly ranked nodes and re-learning low-dimensional node representations for DeepWalk and LINE embedding methods.

Many metropolitan cities motivate people to exploit public bike-sharing programs as alternative transportation for many reasons. Due to its’ popularity, multiple types of research on optimizing public bike-sharing systems is conducted on city-level, neighborhood-level, station-level, or user-level to predict the public bike demand. Previously, the research on the public bike demand prediction primarily focused on discovering a relationship with weather as an external factor that possibly impacted the bike usage or analyzing the bike user trend in one aspect. This work hypothesizes two external factors that are likely to affect public bike demand: weather and air pollution. This study uses a public bike data set, daily temperature, precipitation data, and air condition data to discover the trend of bike usage using multiple machine learning techniques such as Decision Tree, Naïve Bayes, and Random Forest. After conducting the research, each algorithm’s output is evaluated with performance comparisons such as accuracy, precision, or sensitivity. As a result, Random Forest is an efficient classifier for the bike demand prediction by weather and precipitation, and Decision Tree performs best for the bike demand prediction by air pollutants. Also, the three class labelings in the daily bike demand has high specificity, and is easy to trace the trend of the public bike system.

The process of acquiring the large-scale complex systems is usually characterized with cost and schedule overruns. To investigate the causes of this problem, we may view the acquisition of a complex system in several different time scales. At finer time scales, one may study different stages of the acquisition process from the intricate details of the entire systems engineering process to communication between design teams to how individual designers solve problems. At the largest time scale one may consider the acquisition process as series of actions which are, request for bids, bidding and auctioning, contracting, and finally building and deploying the system, without resolving the fine details that occur within each step. In this work, we study the acquisition processes in multiple scales. First, we develop a game-theoretic model for engineering of the systems in the building and deploying stage. We model the interactions among the systems and subsystem engineers as a principal-agent problem. We develop a one-shot shallow systems engineering process and obtain the optimum transfer functions that best incentivize the subsystem engineers to maximize the expected system-level utility. The core of the principal-agent model is the quality function which maps the effort of the agent to the performance (quality) of the system. Therefore, we build the stochastic quality function by modeling the design process as a sequential decision-making problem. Second, we develop and evaluate a model of the acquisition process that accounts for the strategic behavior of different parties. We cast our model in terms of government-funded projects and assume the following steps. First, the government publishes a request for bids. Then, private firms offer their proposals in a bidding process and the winner bidder enters in a con- tract with the government. The contract describes the system requirements and the corresponding monetary transfers for meeting them. The winner firm devotes effort to deliver a system that fulfills the requirements. This can be assumed as a game that the government plays with the bidder firms. We study how different parameters in the acquisition procedure affect the bidders’ behaviors and therefore, the utility of the government. Using reinforcement learning, we seek to learn the optimal policies of involved actors in this game. In particular, we study how the requirements, contract types such as cost-plus and incentive-based contracts, number of bidders, problem complexity, etc., affect the acquisition procedure. Furthermore, we study the bidding strategy of the private firms and how the contract types affect their strategic behavior.

In many complex domains, the input data are often not suited for the typical vector representations used in deep learning models. For example, in knowledge representation, relational learning, and some computer vision tasks, the data are often better represented as graphs or sets. In these cases, a key challenge is to learn a representation function which is invariant to permutations of set or isomorphism of graphs. In order to handle graph isomorphism, this thesis proposes a subgraph pattern neural network with invariance to graph isomorphisms and varying local neighborhood sizes. Our key insight is to incorporate the unavoidable dependencies in the training observations of induced subgraphs into both the input features and the model architecture itself via high-order dependencies, which are still able to take node/edge labels into account and facilitate inductive reasoning. In order to learn permutation-invariant set functions, this thesis shows how the characteristics of an architecture’s computational graph impact its ability to learn in contexts with complex set dependencies, and demonstrate limitations of current methods with respect to one or more of these complexity dimensions. I also propose a new Self-Attention GRU architecture, with a computation graph that is built automatically via self-attention to minimize average interaction path lengths between set elements in the architecture’s computation graph, in order to effectively capture complex dependencies between set elements.Besides the typical set problem, a new problem of representing sets-of-sets (SoS) is proposed. In this problem, multi-level dependence and multi-level permutation invariance need to be handled jointly. To address this, I propose a hierarchical sequence attention framework (HATS) for inductive set-of-sets embeddings, and develop the stochastic optimization and inference methods required for efficient learning.

This research presents thermal image analysis for Fault Detection and Diagnosis (FDD) of Photovoltaic (PV) Systems. The traditional manual approach of PV inspection is generally more time-consuming, more dangerous, and less accurate than the modern approach of PV inspection using Aerial Thermography (AT). Thermal image analysis conducted in this research will contribute to utilizing thermography and UAVs for PV inspection by providing a more accurate and cost-efficient diagnosis of PV faults. In this research, PV module inspection was achieved through two steps: (i) PV monitoring and (ii) PV Fault Detection and Diagnosis (FDD). In the PV monitoring stage, PV cells were monitored by aerial thermography. In this stage, the thermal data was acquired for the next step. In the PV FDD stage, hot spot phenomenon and the condition of the PV modules were detected and measured. The proposed research will help with the problems of the modern PV inspection and, eventually, contribute to the performance of PV power generation.

Trees are common data structures that are used in many programs and applications. In its simplest form, a binary tree can be used to store numbers in sorted manners. Kd-trees, render trees and abstract syntax trees are more sophisticated examples of tree structures. Furthermore, in functional programming algebraic data types are essentially tree structures as well. In several tree-based applications, a tree is constructed, and several traversals traverse the tree to perform different computations. Tree fusion is a transformation that targets combining and fusing different traversals that traverse the same tree and perform them together (ideally in one traversal). Traversal fusion has several performance benefits such as reducing the traversing overhead and the memory accesses, enhancing locality, and eliminating intermediate structures. Previous work has been done on fusion and was mostly successful either in specific domains or limited scopes. This work introduces novel techniques for performing fusion in both imperative and functional programming settings with a focus on generality. The new techniques target general traversals; minimizing the burden on programmers and increasing the coverage of the transformation. Furthermore, it exploits fusion opportunities that previous approaches do not, achieving significant speedups for a wider range of programs.