Traditionally, we use networks to securely and efficiently convey specific information messages to one or more receivers. However, communication networks today are increasingly used to serve a fundamentally different traffic, that delivers types of content rather than specific messages. For instance, when we want to find photos of an event, we may not care which specific photos we receive - we only care about the type of content, namely, that they are photos of the correct event. Content-type traffic pervades a host of applications today, e.g., search engines, recommender systems, and advertising networks. Research on content-type networks is very popular. Most of the existing work looks at how to classify content into types; what to replicate, where and how to store, and from where to retrieve specific data. In contrast, we investigate a totally different question: are there benefits in designing information transmission schemes specifically tailored to content-type traffic?

Our research indicates that in some cases, these benefits can be significant. We design a polynomial-time algorithm for pliable index coding that requires at most $O(\log^2(n))$ broadcast transmissions to serve $n$ clients, as compared to $O(n)$ broadcast transmissions for conventional index coding. This indicates that the exponential benefits of pliable index coding can be effectively realized. Moreover, we explore two applications: recommender systems and distributed computing. In recommender systems, we ask: how much we can gain in terms of bandwidth and user satisfaction, if recommender systems took into account not only the user preferences, but also the fact that they may need to serve these users under bandwidth constraints, as is the case over wireless networks. In other words, what if the recommender systems became bandwidth aware? In this setup, the user is satisfied to receive any message she does not already have, with a satisfaction proportional to her preference for that message. We show, through a number of scenaria, that although the optimization problems are in general NP-hard, polynomial time algorithms with constant approximation ratio can be designed to achieve more than 80\% of the satisfaction and to save 90\% of bandwidth. In distributed computing, to improve the communication efficiency in the data shuffling phase, we examine the pliable index coding problem under data shuffling constraints, where each of the $m$ messages can satisfy at most $c$ out of $n$ clients. We show that the constrained pliable index coding can achieve up to $O(n)$ (best case) benefits over index coding. We prove that the problem is NP-hard and the optimal broadcast transmissions for random instances is almost surely upper bounded by $O(\min\{\frac{n}{c\log(n)},\frac{n}{\log(m)}\})$ for $c=o(\frac{n^{1/7}}{\log^2(n)})$ and $O(\min\{\frac{n}{c}+\log(c),\frac{n}{\log(m)}\})$ for $c=\Omega(\frac{n^{1/7}}{\log^2(n)})$. Building upon constrained pliable index coding, we design a hierarchical data shuffling scheme for distributed computing. By leveraging the many possible shuffling choices, our proposed shuffling scheme is able to reduce the communication cost and achieve benefits up to $O(ns/m)$ over the index coding method, where $ns/m$ is the average number of workers caching a message, and $m$, $n$, and $s$ are the numbers of messages, workers, and cache size, respectively. In addition, we study the beneficial cases of content-type coding over large scale networks and over erasure channels. Compared with message-specific coding, we show that the benefits can be up to the number of messages in the content type for the former case and up to $19.5\%$ for the latter case with symmetric setting.

Over the past few decades, engineering and statistical principles have played a significant role in the advancement of biology and medicine. In this dissertation, we study two problems which have applications in bioinformatics and epidemiology, respectively. We look at these problems through the lens of statistical inference, and propose exact and approximate algorithms.

The first line of work is on sequence reconstruction over deletion channels, which lies at the intersection of theoretical computer science, information theory and bioinformatics. Deletion channels have become increasingly relevant in bioinformatics to model impairments in DNA sequencing and DNA data storage. Prior works in theoretical computer science have proposed sequence alignment heuristics for deletion impairments. We take an alternate approach to this problem and propose algorithms that are rooted in statistical principles, such as maximum-likelihood decoding and maximum-aposteriori decoding. Our algorithms outperform the state-of-the-art algorithms in this space. Along the way, we also develop mathematical tools that are of general interest beyond the study of deletion channels.

The second line of work is on group testing which is related to information theory, medical testing and epidemiology. Group testing is a technique that pools together diagnostic samples from multiple individuals to identify infected individuals in a population. Such a technique potentially allows us to use fewer tests than what is required to test everyone individually. Here again, we improve upon the state-of-the-art group testing algorithms by developing a new statistical inference approach to this problem. Moreover, while traditional group testing assumes independent infections, we observe that contagious diseases like COVID-19 are governed by community spread. We show that taking into account the side information provided by the community structure may lead to significant savings – upto 60% fewer tests compared to traditional test designs in our experiments. We develop new bounds and new approaches to encoding and decoding algorithms that take into account the community structure and integrate group testing into epidemiological modeling.

Fully autonomous vehicles are rapidly approaching realization and concerns regarding their safety and robustness are a prominent obstacle to their integration into society. A bottleneck of an autonomous vehicle’s safety is its ability to self-localize under all types of conditions. Self-localization is critical for the vehicle to route to a destination, for the vehicle itself to be tracked in case of theft, and in determining local traffic laws to operate harmoniously with other traffic and pedestrians. The global positioning system (GPS) is utilized by autonomous vehicles for self-localization, but the availability and reliability of GPS is not guaranteed in all situations, such as when GPS reception is weak or when an adversary is spoofing information. We address this problem by proposing a self-localization method for autonomous vehicles that does not require GPS at localization time. The framework herein describes the representation of a road network as a graph using the data made available by OpenStreetMap, and an encoding of street segments based on detected landmark objects. We derive the current state observer transition function from the resulting graph, and employ localization methods based on set distance and minimum cost paths to determine the most probable location of the vehicle given a sequence of observed landmarks. Through simulating vehicle traveling paths along a road network generated from real data of a region of Washington D.C., we evaluate the performance of our method for varying degrees of landmark observation error and analyze the algorithm’s time and memory complexity, demonstrating that the approach provides a feasible solution to the problem.

With the increase of inter-connected devices, it is of paramount importance to ensure the security of the exchanged information. While cryptographic techniques provide tools to provide confidentiality and data integrity, such techniques may not provide the most efficient solutions. In addition, applications today have challenging performance requirements, with the emergence of time-critical applications such as vehicle networks, as well as resource-limited inter-connected devices such as the devices used in the Internet-of-Things. For such applications, novel security solutions that are application-tailored are needed to meet the performance requirements while adhering to the constraints imposed by the available resources. In this thesis, we adopt this methodology for security design: by understanding the nature of the application, the possible adversaries that may target the communication system, as well as the performance requirements, we design suitable and efficient security solutions. We show this methodology in the context of three different scenarios. The first scenario is data broadcasting in the context of the index coding problem. We study the problem of providing privacy guarantees against curious clients who are interested in knowing the requests and side information sets of other clients. We first design index codes with higher privacy levels than conventional index codes. We also provide a mechanism, which we call k-limited-access schemes, which transforms any index coding technique into another code with higher privacy guarantees. The second scenario is in the context of communication systems which relies on millimeter waves. We tackle the problem of secret key establishment. We propose a secret key establishment protocol which allows two communicating parties to establish shared secret keys at very high rates. We showcase the performance of our proposed technique in two different applications: in millimeter wave wireless systems such as 5G networks and IEEE 802.11ay, and vehicle platooning. The last scenario is in the context of Cyber-Physical Systems. We first argue that, in many situations, an adversary is interested in learning the state vector of the control system. In such cases, a more suitable security metric would be a distortion-based one which leads the adversary to make state estimates that are far from the actual value. We then propose security schemes that require a very small number of secret key bits and still perform well according to the proposed metric: we show that our proposed schemes are in fact optimal for many cases.

Next-generation wireless networks are expected to usher in a revolution of communication systems in terms of the network capabilities, technologies used, and serviced applications. A key aim in these networks is to expand available bandwidth and enable multi-gigabit emerging applications that range from ultra-high-definition video to autonomous vehicle platoons and Industry 4.0 low latency communication. The focus of this dissertation is to study two technologies that are envisioned to drive the physical layer in the next generation of networks - multi-hop relay selection and mmWave transmission - in terms of fundamental theoretical guarantees on performance and the efficiency of operation and scheduling.

We study the problem of relay selection for routing in a wireless full-duplex relay network. We derive fundamental worst-case guarantees on the performance of routing in wireless relay networks operating in full-duplex with arbitrary topology. This extends the state of the art in terms of these fundamental guarantees which were previously only studied for the diamond network topology. These worst-case guarantees provide guidance for heuristic and optimal routing approaches in full-duplex wireless networks.

Second, we study the problem of efficiently selecting a subset of relays in a diamond network operating in half-duplex and the fundamental guarantees that follow from the selection of the best relay subset. We prove fundamental guarantees on the retained capacity when a subset of N-1 relays are retained out of N relays in a diamond topology, For selecting k

Next, we show that selecting the simple route with the highest approximate capacity with a half-duplex network with arbitrary topology is in general an NP-hard problem (unlike its full-duplex counterpart). Additionally, we provide sufficient properties on the network topology that ensure that finding the best route can be done in polynomial-time in the number of nodes in the network. These results are enabled by a closed-form expression that we derive for the approximate capacity of a half-duplex network that enabled the study of its NP-hardness as well as deriving an efficient routing algorithm.

Finally, we study the impact of directional communication which is envisioned in mmWave communication networks. To tackle this, we propose a new information-theoretic model for multi-hop wireless networks referred to as "1-2-1 network". In a 1-2-1 network, at any time instance, two nodes can communicate only if they point beams (which need to be scheduled) at each other, while if they do not point beams at each other, no signal can be exchanged. We used this model to approximate the Shannon capacity for mmWave networks with arbitrary topology operating with full-duplex or half-duplex mmWave nodes. Additionally, we develop a provably optimal polynomial-time algorithm to compute the approximate capacity and an optimal beam-orientation schedule in full-duplex and half-duplex mmWave networks.

In this thesis, we consider secure communication in the presence of an eavesdropper. With the explosion in the growth of the data produced and communicated, sensitive information such as financial transactions, health records, and control signals for cyber-physical systems, has to be securely exchanged. Today, the ever-increasing computational power of adversaries is challenging the state-of-the-art cryptographic encryption mechanisms, as these mechanisms assume adversaries with limited computational power. Thus, with the advent of the quantum computing era, we require new mechanisms to guarantee a secure exchange of information. Moreover, the growing number of small and energy constrained connected devices involved in data exchange calls for lightweight encryption schemes, as low-complexity devices cannot implement complex schemes.

We consider three different scenarios and exploit specific opportunities present in each of these scenarios; we develop lightweight encryption schemes that do not require sharing large keys in advance and are secure against eavesdroppers with unlimited computational capabilities.

The first scenario we consider is multiple unicast traffic over wireline networks. In these networks, a single source is connected to m destinations interested in different messages. In designing encryption schemes, we exploit the fact that although the eavesdropper is computationally super-powerful, it might not have capabilities to eavesdrop the entire network. We use the multi-path diversity to securely communicate against the eavesdropper without requiring any pre-shared key.

The second scenario we consider consists of millimeter wave (mmWave) networks. mmWave communication requires deploying networks of relays that communicate through directional beams to compensate for the high path-loss and the high blockage. Since we need to use beamforming and align beams to activate links, we cannot use all the underlying links of the network simultaneously. However, the degree of freedom in choosing the links to activate can be leveraged for secure communication against an eavesdropper. We show that we can achieve a secure capacity that in some cases, can be very close to the unsecure capacity. Here, capacity refers to the maximum flow of information over the network.

For the third scenario, we consider cyber-physical systems and propose a distortion based security framework where the distortion measures the distance between the eavesdropper's estimates and the ground truth. The primary motivation for this framework is that the messages exchanged in these systems are embedded in a metric space having a notion of distance, and securing raw bits as in traditional encryption schemes might not be necessary. Instead, we show with an example of a linear dynamical system that a carefully designed encryption scheme can significantly distort the eavesdropper's view with just one bit of the pre-shared key.

The growing availability of data and rapid advancements in machine learning are revolutionizing decision-making. Often, these data come from distributed devices with low computational capabilities, connected to a central learner through communication constrained channels. As a result, developing learning algorithms with low communication and computational demands is increasingly attracting attention in the literature and in practice. In this thesis, we solve open problems within this framework for three popular learning setups: Contextual Bandits, Multi-Modal Representation Learning, and Classification.Contextual Bandits is an online learning problem where an agent makes decisions based on contextual information and receives feedback (rewards) for these decisions. The agent learns from past interactions to optimize future rewards. This problem models various practical applications such as recommendation systems, clinical trials, and resource allocations. This thesis offers the following contributions to this area. First, we propose a method that uses only 3 bits to send each reward, even if the rewards take values from an infinite set, while achieving (nearly) optimal learning performance. Second, we prove a surprising reduction from contextual to non-contextual bandits (where only a single context is available for all users). This allows to solve contextual bandits without the need to communicate the context, which can be communication heavy in many practical setups. The reduction provides a framework for developing efficient contextual bandit algorithms by using the simpler algorithms proposed for non-contextual bandits, leading to improved performance bounds in a number of setups such as contextual bandits with sparse unknown parameters, misspecification, privacy constraints, adversarial corruption, and many others. The reduction takes a step into solving multiple open problems in these setups. Lastly, we introduce the first computationally efficient algorithm for contextual bandits with limited policy switches, which is highly relevant to most practical scenarios. In most setups, the learner cannot frequently update the policy due to communication or computation limits, or because of high response rates, as in the case of clinical trials or recommendation systems. In the second setup, Multi-Modal Representation Learning, the focus is in extracting common and private features from multi-modal data, which is required for various applications such as vehicle tracking and medical diagnosis from multi-modal sources, e.g., images, and audio. This thesis introduces the notion of Common Information Dimension (CID), which quantifies continuous common information between sources in terms of number of dimensions. Unlike existing bit-based measures that cannot deal with continuous common information, CID provides a fundamental limit on the dimension of common latent variables in multi-modal learning applications. We develop methods to compute CID for distributions of interest to practical scenarios. Lastly, the Classification setup aims to compress features that are collected at distributed nodes and sent to a central entity for classification. This scenario is encountered by many machine learning applications including wireless cyberphysical systems, immersive environments and supported health. This thesis proposes a compression scheme tailored to classification tasks that allows the central entity to use an existing classifier that operates transparently to the feature compression and maintains the (uncompressed) performance. Although we prove the NP-hardness of finding the optimal compression scheme, we introduce a computationally efficient compression algorithms to approximately solve the problem. We experiment our schemes in practical setups showing significant savings in both the communication and computation costs while achieving state-of-the-art performance.