In summary, the focus of my research group is on coding and information theory and their interplay with communications and learning. The impact of our research includes scientific impact on the one hand and advancement of technology on the other hand. Our research answers fundamental questions that lie at the intersection of statistics, abstract algebra, and combinatorics, while advancing various key elements in communication, storage, and learning systems. This includes reliability, security, and privacy of these systems.



New Faces of Channel Coding in Wireless Communications

Channel codes are essential parts of any wireless communication system in order to provide reliable connectivity. As a result of recent developments in coding theory, there has been an entire reform in channel coding in 5G systems by incorporating low-density parity-check (LDPC) and polar codes into the standard. However, as 5G networks are being unveiled across the globe and discussions about 6G are underway, new channel coding paradigms emerge necessitating fresh investigations efforts in this domain. To this end, we have launched several new studies targeting these emerging problems. This includes code design for heterogeneous wireless systems where coded bits are transmitted across a wide range of frequency bands experiencing uncorrelated channel conditions and coding for channels under extreme conditions for power-constraint Internet-of-Thing (IoT) networks and deep-space satellite communications providing connectivity to remote parts of the globe.



Machine Learning-Assisted Error Correction

It is envisioned that machine learning and artificial intelligence (AI) offer the potential for low-complexity and cost-efficient solutions in future generations of communication and data networks. Motivated by this, we are pursuing two separate objectives in the context of channel coding for future networks. In one direction, we focus on designing neural network-based efficient and low-complexity decoders that provide competitive and near-maximum likelihood performance with efficient implantation that is robust with respect to channel variations. In another direction, we change the code structure as well and invent new non-linear codes using neural network architectures in order to compete with and even to beat state-of-the-art channel codes in various wireless settings especially when coding and modulation are designed jointly over real/complex numbers.



Privacy-Preserving Machine Learning

Data-driven machine learning (ML) has become ubiquitous in the past decade given the abundance of available data. As the number of technologies and applications relying on these advancements continues to grow at an unprecedented scale, the users’ data privacy will be increasingly under attack by adversaries and malicious users. Furthermore, the “right to be forgotten” laws allow users to ask for removal of their data as well as any trace of it from learned models in specific platforms. This gives rise to the following fundamental questions: How to train a machine learning model while keeping the data private? And how to remove the trace of data from the model after the model is trained? To address these critical problems, we develop methods for privacy-preserving learning solutions that work over real-valued data and are scalable with the number of clients. We also work on leveraging coding-theoretic methods toward machine unlearning for efficient removal of data from learned models upon user’s requests.



Subspace Coding, Compression, and Dimensionality Reduction

In subspace codes, the information is embedded into subspaces as linear-algebraic objects in an ambient vector space. Hence, they become relevant whenever the medium, either a communication or a computation platform, preserves the linear span of input vectors rather than the specific vector entries. In this direction, we work on developing new paradigms for reliable communications over networks by utilizing subspace coding, as opposed to conventional block coding. Furthermore, we work on the dual problem in the compression domain, which is central to a wide range of applications involving large-scale raw data often exhibiting low-dimensional structures. More specifically, we work on randomized algorithms and sketching protocols for compression of high-dimensional data, represented in terms of subspaces or matrices, with low-dimensional structures.



Secure and Private Communications

The forecast of tremendous growth of wireless networks in future systems poses a higher risk of malicious attacks against message confidentiality in communication systems. The conventional cryptographic techniques currently deployed in wireless systems are based on point-to-point encryption and decryption protocols. These protocols require a shared secret key that is only known to the legitimate parties. In contrast to conventional cryptographic algorithms, physical layer security methods are keyless, where the noise level of the wireless link is utilized to provide security. Also, no computational restrictions are placed on the adversary, and hence, there is no need for unproven assumptions of computational hardness. However, physical layer security (PLS) methods often rely on arguable assumptions on the quality of eavesdropper’s channel. In order to establish real-world hardware solutions that realize PLS methods in practice, we pursue a new framework by means of coupled dynamical systems. Such systems are already used in practice for synchronization purposes and our approach is to utilize them for implementing PLS in the antenna front end. In another direction, we explore physical layer protocols for covert communications, where even the existence of communication is hidden from an eavesdropper.




Vehicular Communications and Networking

As the vehicular communication technology is becoming increasingly ubiquitous, the amount of data downloaded by vehicles from the network as well as offloaded by them to the network continues to increase exponentially. As a result, the wireless bandwidth usage by vehicular networks continues to increase making such resources increasingly scarce and expensive. At the same time, the abundance of communication mechanisms in vehicular networks, including cellular, Wi-Fi, and Bluetooth, provides a unique opportunity to form vehicular mesh networks for exchanging data locally without congesting the network traffic. To this end, we are studying novel frameworks to utilize the existing vehicular connectivity functionalities, including cellular, WiFi, and Bluetooth, to form mesh networks of locally clustered vehicles where certain vehicles in the cluster will be designated as moving base stations (BSs) in such mesh networks. The other vehicles in this network will be then able to transmit/receive large amounts of data to/from other vehicles as well as sharing emergency messages with other vehicles in the mesh network in order to manage the traffic imposed by vehicular communication on cellular networks.



Research Funded By