Research

cloud My research interests lie at the intersection of communications, signal processing, machine learning, and network information theory, with an emphasis on fifth-generation (5G) and beyond radio access technologies, physical layer security, the Internet of things (IoT), and sensor networks. I am investigating the mathematical limits of communication in wireless networks and applying this understanding to improve the design of 5G and beyond networks. An overall theme in my research is to draw upon mathematical and learning tools to understand the limits of communication in wireless networks, and also to figure out how we can approach these limits in practice. My Ph.D. dissertation combined tools from signal processing and coding theory to develop a new framework for distributed source coding, suitable for low-delay communication. Recently, I have been working on new multiple access techniques that can accommodate a massive number of devices envisioned for IoT and 5G networks, in general. I have been working on beamforming techniques to improve the throughput and security of wireless networks of multi-cell multiple-input multiple-output (MIMO) communications systems.

Research interests

  • Machine Learning/Deep Learning for Communications

  • 3D Wireless Networks/Unmanned Aerial Vehicles (UAVs)

  • Wireless Communications and Networking

  • Signal Processing for Communications

  • Non-Orthogonal Multiple Access (NOMA)

  • Wireless Security (Physical Layer Security)

  • (Network) Information Theory and Coding

  • MIMO Beamforming/Interference Alignment

  • Cognitive and Cooperative Communications

  • Compressed Sensing and Data Compression

Applications

  • Beyond 5G (B5G) cellular systems

  • Internet of things (IoT) networks

  • Wireless sensor networks and smart grid

  • Green and energy-efficient communications

  • Delay-sensitive communications


Summary of Projects
I have worked on several projects on large-scale wireless systems (including mobile and sensor networks) in collaboration with leading companies to design/optimize real-world complex systems. Some of my research projects, together with a statement of their potential impacts, are outlined below:

Wireless System Implementation

To fulfill the need for massive numbers of connections with diverse requirements in terms of latency and throughput, 5G and beyond cellular networks are experiencing a paradigm shift in design philosophy: shifting from orthogonal to non-orthogonal design in the waveform, multiple access, and random-access techniques. Interference cancellation is a key component in non-orthogonal systems. In this experimental work, we have implemented a downlink, two-user non-orthogonal multiple access (NOMA) systems. In NOMA, the user with a stronger channel gain (typically, the user nearer to the base station) employs successive interference cancellation (SIC). The SIC technique allows the stronger user to decode and cancel the interfering signal of the other users and boost up its own rate. While the capacity region is obtained assuming perfect channel state information (CSI), in practice CSI needs to be estimated which affects both the quality of the SIC and uses part of the bandwidth for pilot transmission.

We have implemented the two-user NOMA on an NI USRP-2974 wireless testbed to check its spectral efficiency in practice. Experimental results in a real-world building environment show that, NOMA can improve the spectral efficiency about 1 bit/s/Hz compared to orthogonal multiple access (OMA). Please watch the video and refer to the paper below for more details.

  Over-the-air implementation of NOMA: New experiments and future directions,
    IEEE Access, vol. 9, pp. 135828–135844, October 2021.

Deep Learning for Communications

Machine learning (ML) provides a data-driven approach to learning information and solving traditionally challenging problems without relying on predetermined models and equations. Deep learning (DL) is an emergent subfield of ML that has tremendous potential to be applied to almost every industry and different research areas, thanks to new powerful DL software libraries and specialized hardware. Our group is applying different forms of DL (supervised, unsupervised, and reinforcement learning) to solve various complicated problems in communications.

Modulation Classification and Outlier Detection

In this industrial project (funded by L3Harris), we have designed, trained, and tested neural networks for modulation classification and outlier modulation detection. For the former, we have proposed a convolutional neural network (CNN) that can estimate SNR and use it for improving modulation classification. For the latter, we have applied an autoencoder. Some results are published and more are to come.
Related papers:

 1. Handcrafted and neural network based features for outlier modulation detection,
    IEEE Asilomar Conference, October 2022.

 2. Strategies for enhanced signal modulation classifications under unknown symbol rates and noise conditions,
    submitted to the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.

Precoding and Resource Allocation

Our goal in this project is to come up with new ML/DL-based solutions for a few different but related problems. Specifically, we have developed supervised, unsupervised, and reinforcement learning based precoding and resource allocation as described below.

  • Supervised DL: we have developed supervised learning algorithms for beamforming to enhance PHY security, harvest energy, and transmit reliably, in [C24], [J20], and [J22]. We have shown that a well-trained neural network achieves a near-capacity rate but is much faster than the state-of-the-art solution using much less memory. This approach is promising for IoT applications, which intrinsically have limited computation abilities and battery life.
  • Unsupervised DL: we have designed autoencoder-based end-to-end transmission for MIMO transmission in [C34]. This work is being extended to various other settings, including interference channels [C38].
  • Deep Reinforcement Learning: we have used deep reinforcement learning for clustering and power allocation in NOMA problems in [J17]. We have also introduced using deep reinforcement learning for inter-cell interference mitigation in mmWave multi-cell networks in [C37].

    Related papers:

     1. Power allocations in cache-aided NOMA systems: Optimization and deep learning approaches
        IEEE Transactions on Communications, vol. 68, no. 1, pp. 630–644, January 2020.

     2. Deep learning based precoding for the MIMO Gaussian wiretap channel
        Global Communications Conference (GLOBECOM) Workshops, Waikoloa, HI, December 2019.

     3. Multi-objective DNN-based precoder for MIMO communications,
        IEEE Transactions on Communications, vol. 69, no. 7, pp. 4476–44880, July 2021.

     4. Secure precoding in MIMO-NOMA: A deep learning approach,
        IEEE Wireless Communications Letters, vol. 11, no. 1, pp. 77–80, January 2022.

     5. Deep reinforcement learning for interference management in millimeter-wave networks,
        Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, California, October 2022.

     5. SVD-embedded deep autoencoder for MIMO communications,
        IEEE International Conference on Communications (ICC), Seoul, South Korea, May 2022.

    Multiple Access Techniques for Massive Connectivity

    Future radio access networks are expected to support a massive number of users with a diverse set of requirements in terms of delay and throughput. These requirements challenge different aspects of the current cellular networks including the multiple access (MA) methods [B2]. A common feature of newly designed MA schemes is the use of non-orthogonal multiple access (NOMA) schemes instead of conventional orthogonal schemes. NOMA is a potential enabler for the development of 5G wireless networks [J15]. Compared to the conventional orthogonal multiple access schemes such as TDMA and FDMA, NOMA can scale up the number of served users, increase spectral efficiency, and improve user-fairness. As such, NOMA has been proposed for the 3rd Generation Partnership Project (3GPP) Long Term Evolution–Advanced (LTE–A) standards and is envisioned to be a part of 5G cellular networks. Despite significant attention from the communications community, NOMA has been subject to several widespread misunderstandings. In [J13], we identify such common myths and clarify why they are not true. In addition, much of the literature is on single-cell, and much less attention has been given to multi-cell NOMA.

    Our goal in this project is to realistically and thoroughly identify the potentials of NOMA and investigate the potentials and challenges of NOMA in a multi-cell environment and harvest the benefits of NOMA. Inter-cell interference, particularly at the cell edge, is by far the main challenge in multi-cell networks. This interference situation becomes worse when NOMA is used, as cell-edge users constantly experience interference from the neighboring cell, whereas in the case of TDMA/FDMA interference is limited to certain time slots or frequency bands. To deal with this problem in a MIMO communication network, we have developed two new interfere alignment (IA)-based techniques in [J6]. This approach is extended to an arbitrary number of cells in [C16], where the maximum number of users supported by the proposed scheme in multi-cell MIMO networks is characterized too.

    xyz

    Caption: An illustration of multi-cell NOMA networks.

    Related papers:

     1. Non-orthogonal multiple access in multi-cell networks: Theory, performance, and practical challenges
        Springer, 2019

     2. Non-orthogonal multiple access: Common myths and critical questions
        IEEE Wireless Communications, October 2019

     3. NOMA: An Information Theoretic Perspective
        IEEE Wireless Communications, October 2019

     4. Interplay between NOMA and other emerging technologies: A survey
        IEEE Transactions on Cognitive Communications and Networking, December 2019

     5. Non-orthogonal multiple access in multi-cell networks: Theory, performance, and practical challenges
        IEEE Communications Magazine, accepted in October 2017

     6. Coordinated beamformig for multi-cell MIMO-NOMA
        IEEE Communications Letters, January 2017

     7. On the number of users served in MIMO-NOMA cellular networks
        International Symposium on Wireless Communication Systems (ISWCS'16), September 2016


    Physical Layer Security in Multiple-Antenna Networks

    Due to the broadcast nature of wireless channels, wireless communication, such as cellular and WiFi systems, is vulnerable to eavesdropping and malicious attacks. As such, wireless security has been an important concern for many years. As a means of augmenting wireless security, physical layer security has attracted significant attention recently. Physical layer security is based on the information-theoretic secrecy that can be provided by physical communication channels, an idea that was first proposed by Wyner in the context of the wiretap channel. In this channel, a transmitter (Alice) wishes to transmit information to a legitimate receiver (Bob) while keeping the information secure from an eavesdropper (Eve). Wyner proved that it is possible to have both reliable and secure communication between Alice and Bob.

    xyz

    Caption: MIMO Gaussian wiretap channel with nt, nr, and ne antennas at the transmitter (Alice), legitimate receiver (Bob), and eavesdropper (Eve).

    In this project, we have developed optimal precoding and power allocation schemes to reliably transmit secure data over MIMO Gaussian wiretap channels in which Eve and Bob have arbitrary numbers of antennas, but Alice has two antennas [J8]. The basic principle is to make the channel of the legitimate receiver (Bob) stronger, in some sense, than that of the eavesdropper (Eve), exploiting linear beamforming techniques. As an extension of this work, in [J18] we have introduced a new precoding and power allocation method (called rotation-based precoding) for arbitrary numbers of antennas at each node. This method gives a new framework for precoding in MIMO channels and can be applied to various similar problems such as [J14].

    Related papers:

     1. Optimal beamforming for Gaussian MIMO wiretap channels with two transmit antennas
        IEEE Transactions on Wireless Communications, accepted in July 2017

     2. A rotation-based method for precoding in Gaussian MIMOME
        IEEE Transactions on Wireless Communications, in review

     3. A rotation-based precoding for MIMO broadcast channels with integrated services
        IEEE Signal Processing Letters, November 2019

     4. On the secrecy capacity of the Z-interference channel
        International Zurich Seminar on Communications (IZS), Zurich, Switzerland, March 2016


    Interference Management and Cognitive Radio Networks

    Interference management is a central issue in wireless networks. This problem becomes more important as networks get denser and denser. Our understanding of optimal encoding in such channels is limited. Two basic approaches are to orthogonalize users’ signal into different bands and to share the spectrum fully. The former is known as time/frequency division multiplexing, and the latter is treating interference as noise. Both of these approaches are, in general, suboptimal. It is known that Han-Kobayashi (HK) can strictly improve the above approaches. However, the complexity of the HK and its optimal power allocation is unknown. In [C15], we have shown that time-sharing in two subbands is enough to achieve the border of the HK region for the one-sided and a large part of the mixed interference cases. This simple scheme significantly reduces the complexity of the HK inner bound and is the best-known scheme for the above interference cases. The problem of finding the capacity region of the interference channel is notoriously hard but very important as it can give intuitions on optimal signaling for interference networks. Likewise, the capacity region of the cognitive interference channel is an important open problem in network information theory.

    Our objective in this project was to characterize the fundamental limits of communication over the cognitive channel, a building block of cognitive radio networks. In particular, we have studied the capacity region of the cognitive interference channels and cognitive Z-channels. The main goal of this study was to develop coding schemes for cognitive radio technology that enable the cognitive user to make opportunistic use of the spectrum in coexistence with incumbent users. We have characterized the largest capacity region for the two-user cognitive channels in [C6]. Different classes of discrete memoryless cognitive interference channels (DM-CIC) are depicted in the figure below.

    xyz

    Caption: The class of the discrete memoryless cognitive interference channels (DM-CIC).

    Related papers:

     1. The capacity of more capable cognitive interference channels
        Allerton Conference on Communication, Control, and Computing (Allerton), October 2014

     2. Comments on new inner and outer bounds for the memoryless cognitive interference channel and some new capacity results
        IEEE Transactions on Information Theory, June 2013.

     3. The capacity of less noisy cognitive interference channels
        Allerton Conference on Communication, Control, and Computing (Allerton), October 2012

     4. On the capacity of the cognitive Z-interference channel
        Canadian Workshop on Information Theory (CWIT), May 2011


    Transmission Strategies for Sensor Networks

    Sensor networks require energy-limited communication/computing schemes, as sensors usually run on batteries, and it is required that the battery lasts for a long time. Distributed source coding (DSC) is a key enabling technology to this end. It refers to the compression of multiple, statistically dependent sources (e.g., neighboring sensors outputs) that do not communicate with each other and thus are encoded in a distributed manner. DSC has been privileged by the advancement of capacity-approaching channel codes, e.g., turbo and low-density parity-check (LDPC) codes. However, the performance of turbo/LDPC-based DSC systems deteriorates largely when the code length reduces. In practice, where delay and complexity limitations are stringent, code length cannot be very large. Thus, other low-delay and energy-limited communication strategies are crucial for sensor networks.

    xyz
    Caption: Distributed source coding based on real-number codes

    With this in mind, in a project sponsored by Hydro-Quebec and aimed at information transmission in substations, we have developed a new framework for distributed lossy source coding with application in low-delay sensor networks. To this end, we have proposed a scheme in which real-number, rather than binary, codes are used to compress correlated distributed sources. In other words, we first compress the continuous-valued sources and then quantize them, as opposed to the conventional approaches wherein compression is done after quantization. The advantage of this method is twofold. First, the correlation channel can be modeled more realistically, as it captures the dependency between the continuous-valued sources rather than quantized ones. Further, by using discrete Fourier transform (DFT) codes for binning (compression), we are capable of reducing the quantization loss by a factor of code rate. The proposed technique is specifically appropriate for delay-sensitive sensor networks, as it performs well with short codes.

    In this project, we have developed