Abstract

As the largest online encyclopedia, Wikipedia provides a large number of human-edited articles that are linked to named entities. For growing the Wikipedia content, determining whether such entities require their own Wikipedia article and then producing these articles by hand is a laborious task. The limited number of active contributors and the huge time spent in collecting the information and editing is making this manual process slow and difficult. Our thesis presents an automated solution for both determining whether a named-entity warrants its own article in Wikipedia and thereby creating one using a semi-supervised approach. Wikipedia requires a keen investigation about which articles to be included for it to maintain its indispensability. To prevent unnecessary articles from being included, official guidelines of Wikipedia demand these named entities meet “notability” standards for their article inclusion. We investigated automation around notability for named entities by using reliability and entity salience features. Evaluations of our notability determination system provide evidence that our solution can replace the manual decisions made by the reviewers for the inclusion of an article using the notability rules. We also developed a semi-supervised automatic generation of Wikipedia articles as an alternative to its manual creation. A framework was created to generate a Wikipedia article for a named entity, which not only looks similar to other Wikipedia articles in its category but also aggregates the information about diverse aspects related to that named entity from the Web. In particular, a semi-supervised method is used for determining the headings and identifying the content for each heading in the Wikipedia article generated. Additional important contributions here include reducing the manual effort in collecting relevant training data by extracting it from the wiki-dump and increasing the reliability of the content by introducing parameters modeled on the reliability of a site based on existing Wikipedia articles belonging to the same category. Our experiments showed the viability of our solution as an alternative to the previous approaches that used supervised or unsupervised methods. Such an automation for both notability and article creation in Wikipedia can help in paving a path for a future where maintaining consistent growth in the Wikipedia is automatic and requires very little human-interference.

Abstract

The field of computer vision is rapidly expanding and has significantly more processing power and memory today, than in previous decades. Video has become one of the most popular visual media for communication and entertainment. In particular, automatic analysis and understanding the content of a video is one of the long-standing goals of computer vision. One of the fundamental problems is to model the appearance and behavior of the objects in videos. Such models mainly depend on the problem definition. Typically, in many scenarios, the change in problem statement is followed by the changes in the annotation and its complexities. Creating large-scale datasets in this scenario using the manual annotation process is monotonous, time-consuming and non-scalable. In order to address this challenge and strive towards practical large scale annotated video datasets, we investigate methods to autonomously learn and adapt object models using temporal information in videos. Even though the vision community has advanced in field of problem solving but data generation and annotation is still a tough problem. Data annotation is expensive, tedious and involves a lot of human efforts. Even after data annotation, it is essential to validate the goodness of annotations, which again is a tiresome process. To address this problem, we investigate methods to autonomously learn and adapt the object models using temporal information in videos. This involves learning robust representations of the video. The aim of this thesis is two-fold, first we propose solutions for efficient and accurate object annotation mechanisms in video sequences and secondly, to raise awareness in the community about the importance and attention it deserves. As our first contribution, we propose an efficient, scalable and accurate object bounding box annotation method for large scale complex video datasets. We focus on minimizing the annotation efforts simultaneously increasing the annotation propagation accuracy to get a precise and tight bounding box around object of interest. Using a self training approach, we propose a combination of semi-automatic initialization method with an energy minimization framework to propagate the annotations. Using an energy minimization system for segmentation gives accurate and tight bounding boxes around the object. We have quantitatively and qualitatively validated the results on publicly available datasets. In the second half, we propose annotation scheme for human pose in video sequences. The proposed model is based on a fully-automatic initialization, from any generic state-of-the-art method. But the initialization is prone to error due to the challenges in video data type. We exploit the availability of redundant information from the redundant data type. The model is build on the temporal smoothness assumption in videos. We formulate the problem as a sequence-to-sequence learning problem, the architecture uses Long Short Term Memory encoder-decoder model to encode the temporal context and annotate the pose. We show results on state-of-the-art datasets.

Abstract

The objective of a language identification system is to identify the language in a given spoken utterance. Based on the availability of corresponding text to a given utterance, the approach for language identification can either be implicit or explicit. In an implicit case, we only have access to the acoustic signal and statistical models are built over the features extracted from the acoustic signal. In an explicit case, corresponding text is available for the given utterance and rule-based approaches are used to identify the language. This thesis presents an implicit approach based on deep neural network (DNN) models for language identification. DNN models have been recently proposed for the task of language identification. In this thesis, DNN models have been explored in the Indian scenario. While DNN based approach is inherently a frame based one, we propose an attention mechanism based DNN architecture for utterance level classification there by efficiently making use of the context at model level. This approach can be viewed as an alternative to the state-of-the-art Gaussian mixture model based iVector baseline system. On contrary to previously published works using DNNs in LID, DNN equipped with attention mechanism has the advantage of discriminative end-to-end training without any generative component or post-averaging of posterior probabilities obtained at frame level. Several experiments have been performed on a dataset consisting 13 of the official Indian languages with 120 hours of training data. Evaluation of models were performed on 30 hours of testing data with about 2.5 hours for each language. Equal error rate (EER) per language is used as the performance metric. From the results, it is found that the DNN with attention mechanism outperforms the regular DNN and i-vector baseline system indicating the effectiveness of the attention mechanism. Further, a combination of excitation source and vocal tract system features have been explored for the task of language identification. A better performance gain has been achieved by combining the features using a late fusion mechanism. This result indicates the complementary nature of the excitation source information to that of the widely used vocal tract system information.

Abstract

Knowledge of the structure and dynamics of biomolecules is essential for elucidating the underlying mechanisms of biological processes. Given the stochastic nature of many biological processes, like protein unfolding, it’s almost impossible that two independent simulations will generate the exact same sequence of events, which makes direct analysis of simulations difficult. Statistical models like Markov Chains, transition networks etc. help in shedding some light on the mechanistic nature of such processes by predicting long-time dynamics of these systems from short simulations. However, such methods fall short in analyzing trajectories which are oblivious to temporal information, for example, Monte Carlo simulation methods. In this thesis we propose a probabilistic algorithm, borrowing concepts from graph theory and statistical mixture models, to extract folding pathways from molecular dynamic trajectories. A suitable vector representation was chosen to represent each frame in the macromolecular trajectory (as a series of interaction and conformational energies) and dimensionality reduction was performed using principal component analysis (PCA). The trajectory was then clustered using a density-based clustering algorithm, where each cluster represents a metastable state on the potential energy surface (PES) of the biomolecule under study. A graph was created with these clusters as nodes and the most probable path of (un)folding from an initial folded state to a final unfolded state is conceived as the widest path along this graph. We have tested our method on RNA hairpin unfolding trajectory in aqueous urea solution. Our method makes the understanding of the mechanism of unfolding in RNA hairpin molecule more tractable. As this method doesn’t rely on temporal data it can be used to analyze trajectories from Monte Carlo sampling techniques and replica exchange molecular dynamics (REMD). This work is an effort in the direction towards harnessing the full potential of enhanced sampling methods, which until now, is mostly limited to evaluating equilibrium properties of the system

Abstract

More than a decade ago, startups, companies, research groups or individuals had to purchase hardware and set it up for each task. This approach made sense when the utility and the requirement of the hardware were fixed. However, this does not make sense in today’s growing market, since there is a continuous increase or decrease in hardware’s demands. Often, people require large amounts of resources for short duration of time (like in the case of scientific computation). In other cases, people require the ability to scale up and down the number of systems based on the demand rendering dedicated systems unprofitable. Resources can be of different types such as computation(CPU), storage(distributed file system storage), network, databases, etc. Cloud computing opens up a whole new paradigm where individuals or companies can rent all the resources they need and dispose them off later after the work is done. In this process, they are charged for exactly what they use and not a penny more. This is therefore a highly flexible model, since one can add or reduce the resources on the fly according to the requirements and be charged accordingly. With data centers spread across the globe and the use of cloud computing, we can now localize resources closer to the end user to improve performance and experience. Cloud Computing is thus highly profitable as it is based on utilization of commodity hardware. The systems are heterogeneous in nature as they have different memory performance, different storage capabilities, different network capabilities etc. In today’s world, we cannot have homogeneous systems as they will not be able to provide the flexibility that we require. For example, if some nodes in the cluster are over-utilized and some disks in the storage cluster are under-utilized, there can be a lot of scenarios where homogeneous systems won’t work due to the dynamic nature of our workloads or tasks. We therefore need to develop systems that are cognizant of this inherent heterogeneity and can adapt to the constantly changing environment. In our work, we are focused on allocating resources for workloads which are short lived and mainly used for batch processing. During a computing process, workload is the amount of processing that the computer has been given to do at a given time. So our main focus is to allocate resources from a specific cloud provider for handling these dynamic workloads. These assignments could be on the basis of price, utilization, individual cloud characteristics, efficiency/performance etc. We try to use all these factors to design various techniques for the assignment of different types of heterogeneous scenarios. Below are the three scenarios of heterogeneous cloud systems for which we have designed frameworks/algorithms to assign workloads: • In the first scenario, we propose MultiStack, which is a big data orchestration platform for deploying big data jobs across the hybrid cloud environment. The specific architecture elaborated in this paper uses Amazon Web Services as the public cloud provider and Openstack as the private cloud framework. The proposed framework aims at reducing the job completion time of workloads along with decreasing the cost using Spot Instance provisioning instead of on-demand provisioning. In this case, we try to assign an optimal number of instances (cluster of computing resources) to the given workload along with cost minimization. • In the second scenario, we propose a solution for optimal IO Workload assignment with respect to backend storage (Software Defined Storage) using statistical modelling to estimate measures of performance such as Throughput, IOPS, et al in that storage system. The proposed system uses support vector regression to estimate the performance of individual IO Workloads on each available Software Defined Storage(SDS) system for optimal assignment. We demonstrate our solution in a heterogeneous environment comprising of HDFS, GlusterFS, and Ceph. Therefore, in this case, we try to increase the efficiency by choosing/assigning an optimal storage environment for any given type of workload. • In the third scenario, we try to provide optimal assignment of IO workload with respect to compute and storage for a multi-cloud heterogeneous systems. We try to achieve and represent “optimality” by designing a pipeline model which selects a compute and a storage class for a given workload to maximize efficiency and then apply discount schemes provided by the public cloud providers for cost minimization. We also use some input parameters provided by the user to fine tune the cost minimization procedure. We have designed these methods and algorithms using the following three public cloud providers: Amazon Web Services, Microsoft Azure and Google Cloud Platform. We expand on these different scenarios and try to build different methods to solve each of them. Through this work, we hope to help individuals, startups, researc

Abstract

Code-Mixing (CM), a progeny of multilingualism, is defined as a phenomenon where linguistic units such as phrases, words and morphemes of one language are embedded within an utterance of another language. This phenomenon is often observed in conversations within the bilingual and multilingual user group. Also, Code-Mixed language is extensively used on social media sites like Facebook and Twitter. Though code-mixing is the most natural form of conversation both in speech and text (online chats), the current dialog systems and search engines are not capable of handling this kind of social interaction. Many Popular virtual personal assistants used by a significant amount of smart phone users are unable to handle the mixing and switching between any two languages, which occurs very naturally to any bilingual or multilingual person. So, as mentioned above, it becomes extremely important to understand CM for both information extraction and for the purpose of building dialog systems that are capable of social interaction. In this thesis we take the first step towards understanding CM between two languages - Telugu and English, belonging to two different language families, sharing no ancestry at all. We start by building basic preprocessing tools like Language Identification (LID) models, POS taggers for CM data. We collected Telugu-English code-mixed social-media blog based data. The collected data was annotated by two individual annotators and a Kappa score of 88.91 is reported, which is a sign of reliable data. We experimented with various machine learning algorithms and finally we present a Multiple Layer Perceptron (Neural networks) LID model that uses character n-gram vectors augmented with other handcrafted features. Our system gives an F1-score of 97 for English and 96 for Telugu. The LID module has been used for the task of POS tagging of Telugu-English CM data, which produced an accuracy of 52.37%. In the second part of the thesis we made an attempt towards understanding code-mixing in dialog (richest and most natural form of language) through automatic recognition of dialog act (speaker’s intention) of an utterance. We have experimented with learning algorithms like Support Vector Machines, Naive Bayes, Kth Nearest Neighbor and Hidden Markov Models. Our best system that gives an F1-score of 72.30 is HMM based. Non-availability of annotated conversational code-mixed data poses problems and hinders us from adopting data-driven methods like Neural networks. Manual procurement and annotation of code-mixed data is not only time-consuming but also is labor intensive. Therefore, we also investigate on how knowledge extracted from resource rich languages could be useful in dialog act recognition of code-mixed conversations. We show that a decent accuracy of 66% can be obtained by transforming the code-mixed utterances into a sequence of English words by the use of modules like LID, transliteration and translation. We propose that in the future, a variant of transfer learning can be used to incorporate this knowledge along with the knowledge obtained from manually annotated code-mixed conversations.

Abstract

In-band full-duplex (FD) communication that enables concurrent transmission and reception in a single frequency-time channel holds the promise of nearly doubling the spectral efficiency compared to the current half-duplex and out-of-band FD communication systems. The main hurdle in the practical implementation of such a full-duplex system is the loopback self-interference (LSI), which arises due to the FD operation of the communicating nodes. In such an FD transceiver, the transmit signal generally overpowers the intended receive signal making its detection extremely difficult. However, with the recent advancements in the LSI cancellation techniques, FD systems have finally become viable. Hence, it is no wonder that FD has been envisaged as one of the promising technologies for the next generation wireless networks, viz., 5G and beyond. In practice, the LSI cancellation process cannot eliminate the LSI completely due to various reasons, chief among them the difficulty in obtaining the perfect channel state information of the loopback channel. As a result, we are still left with a residual LSI, which would degrade the performance of an FD system. In this work, we consider the design of optimal transceiver and relay processing algorithms for FD multiple-input multiple-output (MIMO) cooperative networks. Specifically, we propose the algorithms for FD communication in three different scenarios: (i) a cognitive radio network, (ii) secure communication using physical layer security techniques, and (iii) a cellular network. In the cognitive radio setup, we consider the FD transceivers and the FD relay to be a part of the secondary network. The transceiver and the relay precoders are designed so as to ensure that the interference caused to the primary user is always below a specified threshold. We have considered the maximization of the receive signal-to-noise ratio (SNR) as the criterion for the design of the transceiver precoders. We have further extended this problem to the case where multiple relays are available for selection. A central control unit selects the relay that provides the highest receive SNR. In the next scenario, a passive eavesdropper tries to decode the data of the two legitimate FD transceivers. In this case, we consider secure communication between the legitimate transceivers using physical layer security techniques. The precoder and receive filter matrices are designed to ensure secure communication between the legitimate users. The secrecy performance of this system in the presence of a passive eavesdropper is analyzed by considering the relay signal as artificial noise for the eavesdropper. The effectiveness of the proposed design is evaluated in terms of the sum-secrecy-rate of the system. Finally, we consider a futuristic FD cellular network scenario where an FD base station communicates with two FD transceiver nodes, where one of the transceivers is assisted by an FD amplify-and-forward relay. In each of the scenarios, our prime focus is on mitigating the residual LSI and its detrimental effect on the performance of the system. The precoders at the transceiver nodes are designed to mitigate the LSI based on the requirements of their operating environment. The LSI at the relay is mitigated by incorporating it into the sum of meansquareerror (SMSE) minimization problem. The relay precoder and the transceiver receive filters are designed based on the minimization of the SMSE of the end-to-end communication. The resulting optimization problem is non-convex and hence a globally optimal solution is not guaranteed. We apply the coordinate descent method and the Karush-Kuhn-Tucker (KKT) conditions to obtain sub-optimal solutions. The complexities of the proposed algorithms are lower than that of the state-of-the-art algorithms, in terms of both number of computations and memory requirements. The various simulation results demonstrate the efficacy of our proposed designs in terms of LSI mitigation, power efficiency, number of iterations, and sum-rate performance.

Abstract

Wireless communication has expanded its horizon in almost every aspect of human lives and its consequence is rapid surge in the demand of spectral resource. Electromagnetic spectrum is a gift of nature and such spectrum bands cannot be increased beyond certain limit. Presently, spectrum is licensed by government agencies to the interested parties by accepting a license fee. At the time of allocation, it is assumed that the licensed party will be utilizing the spectrum up-to its full potential i.e. 100% of the time. Hence the frequency bands to be utilized by the primary user are decided a-priori irrespective of its dynamic utilization status. On the other side with the advent of technologies like IoT, wearable consumer electronics and smartphones etc. applications are becoming bandwidth hungry and the number of connected devices are increasing drastically. Gardner technology research, a premier IoT research company claims that in the year 2015, 4.9 billion devices were connected to the internet and by 2020 this figure will reach to 20.8 billion. Hence with such an exponential increase in the number of connected devices, we will soon be facing scarcity of the electromagnetic spectrum in near future. Cognitive Radio is an innovative technology with unprecedented level of intelligence and capability to enhance the system capacity as well as spectrum agility of wireless communication systems. It mitigates the problem of spectrum scarcity with the aid of opportunistic access of spectrum. This work is dedicated towards developing hardware efficient and reconfigurable spectrum sensing architectures as well as techniques from VLSI design and implementation perspectives. We propose enhancements to the widely used energy detector design which helps in reducing the overall hardware resources (like adders and multipliers) consumed by 46% and alleviates the memory footprint by more than 95%. Further the conventional energy detector design is extended for wide band sensing applications without compromising on detector performance. We also present reconfigurable VLSI architectures for cyclostationary spectrum sensor which aids in sensing OFDM signals constructed with a variety of sub-carrier lengths. We suggested a novel selective sampling based reconfigurable cyclostationaryspectrum sensor which reduces the memory requirement by 99% and other resources by 33%. To evaluate the performance of these architectures, extensive computer simulations are carried out and detector architectures are implemented on an FPGA platform to verify the timing and functionality. Such verification has been carried out by feeding real-world RF signals captured from standard-RF SDR front-ends. Finally, ASIC syntheses and post-layout simulations of these architectures in 90 and 65 nm-CMOS technologies are carried out and its implementation results are compared with the stateof-the-art.

Abstract

An important component of software reliability is the assurance of certain security guarantees, such as absence of low-level bugs that may result in code exploitation, for example. A program crash is an early indicator of possible errors in the program like memory corruption, access violation or division by zero. In particular, a crash may indicate the presence of safety or security critical errors. A safety-error crash does not result in any exploitable condition, whereas a security-error crash allows an attacker to exploit a vulnerability. However, distinguishing one from the other is a non-trivial task. This exacerbates the problem in cases where we get hundreds of crashes and programmers have to make choices which crash to patch first! In this work, we present a technique to identify security critical crashes by applying machine learning on a set of features derived from core-dump files and runtime information obtained from hardware assisted monitoring such as the last branch record (LBR) register. We implement the proposed technique in a prototype called Exniffer. Our empirical results, obtained by experimenting Exniffer on several crashes on real-world applications show that proposed technique is able to classify a given crash as exploitable or not-exploitable with high accuracy.

Abstract

World Wide Web was brought to the fore in 1989 as the first version of HTML (HyperText Markup Language), accompanied with the capability of hypertext - a base format for publishing web pages. As the user types in a website’s address in the address bar of the browser, an HTTP request is initiated by the browser. The browser receives the response from the web server and builds a DOM (Document Object Model), a data structure for storing information in a tree format. It then renders the content from the DOM into its window. Further, Image tag was proposed in 1993, which enabled web pages to cascade requests thereby escalating security threats. The basic browser model, however, allows content inclusion from any server, and this prevails as a basic design flaw which has led to several malicious attacks. Existing browser security policies like Same Origin Policy (SOP), Content Security Policy (CSP) do not mitigate these attacks by design. On March 27, 2015, Github witnessed a massive DDoS attack, the largest in Github’s history till date which prevailed for over three days. In this incident, browsers and users were used as vectors to launch the attack. This class of attack has affected many other websites on a large scale, such as Incapsula and CloudFlare. Upon detailed analysis, we found the primary cause of these attacks to be cross-origin request attacks (CORA). Given the fact that current browser security model does not prevent content inclusion cross-origin attacks, it is rather easy to launch such attacks. CORA is one of the top attacks in web security with CSRF (cross-site request forgery) attack ranking in the top 10 attacks (2017), by OWASP. Against this compelling backdrop, this work undertakes a detailed analysis of CORP, a browser security policy which can be used to mitigate cross-origin request attacks. CORP enables a server to control cross-origin interactions initiated by a browser. The browser intercepts the cross-origin requests and blocks the unwanted requests by the server. Further, we explain the outcome of this work in three folds: Firstly, it demonstrates a compelling application of CORP in mitigating the browser-based DDoS attack, which has severely affected several websites in the past. Secondly, to gain a better understanding of CORP, we built a formal model of cross-origin request attacks and mitigated it using CORP. CORP has a client side dependency where users will have to use a CORP compliant browser to mitigate cross-origin attacks. Considering this, the concluding task was - implementation of CORP in Chromium source code. To evaluate the effectiveness of CORP, we conducted several experiments on various cross-origin attacks. We simulated a browser-based DDoS attack in a lab environment and did a detailed analysis on mitigation of the attack using CORP. After that, we modelled a complex case of a cross-origin transaction, FIM (Federated Identity Management), and incorporated a malicious cross-origin attack in it. We, then, validated CORP by mitigating the attack successfully using CORP. Modelling helped us in identifying the different parts involved in a CORP transaction. Thereon, to gain deeper insight into the functioning of CORP, we implemented CORP in Chromium browser and demonstrated the mitigation of various cross-origin attacks (CSRF, Clickjacking, cross-site timing attacks), present in websites hosted by top enterprises (such as facebook, google, twitter, dropbox, etc.), using CORP.