top of page

The rapid evolution of wireless networks requires stronger guarantees in terms of robustness, adaptability, efficiency, and security than those that are currently available. In my research, I am building novel solutions for emerging and future wireless networks, with an emphasis on spectrum sharing. I dedicate keen attention to incorporating intelligence into my solutions through data-driven machine learning (ML) approaches, which are crucial for robustness and adaptability. A significant focus of my research is to develop realistic systems that go beyond conceptual solutions and can be used in practice. My research has appeared in a diverse set of high-quality publication venues and journals on wireless systems and networks. These include ACM MobiCom, ACM MobiHoc, ACM IMWUT, ACM TOSN, IEEE SECON, IEEE MASS, IEEE DySPAN, IEEE TWC, IEEE TMC. In the following, I categorize my research based on broad themes and briefly summarize my contributions in each of these categories.

Radar Signal Detection

  • Deep Learning-based ESC Sensor for CBRS (ACM MobiCom '21): The FCC has announced a set of rules, known as CBRS, for spectrum sharing in the 3.5 GHz band. A key enabler for CBRS is the ESC sensor, whose task is to detect the presence of radar signals in a timely fashion. We developed DeepRadar, a novel deep-learning-based ESC for detecting radar and estimating their spectral occupancy. The efficiency of DeepRadar lies in its capability of detecting radar signals and estimating their spectral occupancy jointly by transforming the problem into an object detection task and solving it using a customized version of the YOLO algorithm. We implemented DeepRadar using AIR-T software-defined radio (SDR). Our implementation meets all the requirements necessary for ESC certification. Nokia has applied for a patent based on DeepRadar.

  • Detection and Parameter Estimation of Low Power Radar Signals in Interference (to be submitted to IEEE TCCN): Revisited the problem of radar detection and parameter estimation in CBRS, but with a more challenging goal of allowing more interference and detecting lower power radar. We developed a supervised deep learning-based spectrum sensing approach called RadYOLOLet that comprises of two different convolutional neural networks (CNN), one operates on spectrograms and the other on wavelet transforms. RadYOLOLet can achieve 100% radar detection accuracy up to 16 dB SNR. RadYOLOLet can also function accurately under interference up to 16 dB SINR.

  • ML-assisted Radar Signal Processing (IEEE PAST '22): We developed an ML-assisted MIMO radar signal processing pipeline for improved interference estimation, interference suppression, and target tracking under resource constraints. The primary goal of this project was to develop low-complexity and adaptive methods that can leverage GPUs and reap the benefits of larger arrays. The low-complexity aspect is crucial as it enables real-time target tracking. Simulation results show that our ML-assisted interference estimation and suppression can outperform traditional signal processing methods. Our algorithms have been deployed on real radar hardware by Raytheon and Nvidia.

Crowdsourced RF Localization

  • Simultaneous Transmitter Localization for Spectrum Monitoring (ACM MobiCom '17): We developed SPLOT, a novel approach for the localization of simultaneous transmitters for crowdsourced spectrum monitoring. To address the multi-source localization problem, SPLOT uses a simple but efficient method that converts the problem to a set of single transmitter localization problems. Then it uses a matrix inversion approach based on the radio wave propagation path loss model to find the locations of each transmitter. We implemented SPLOT using RTL-SDRs and performed multi-transmitter localization in various experiments. Our experiments show that SPLOT localizes the transmitters with high accuracy in a timely manner.

  • Distributed Measurement Platform for Crowdsensing (IEEE MASS '19, IEEE TMC '23): While experimenting with SPLOT, we realized that crowdsourcing experiments using RTL-SDRs are not scalable and inconvenient for participants. Hence, we designed a novel distributed system for crowdsourced spectrum measurements. The main components that we developed for this system are: a cloud-based server, an untethered, low-power, and portable SDR, called Sitara, and an Android app. The cloud server controls a set of Sitara SDRs, acting as crowdsensing devices. However, since the Sitara SDRs were designed to be untethered, we use the Android app to act as an intermediary between the Bluetooth-enabled SDRs and the cloud server.

  • Learning-based Localization using Crowdsourcing (ACM MobiHoc '20): We developed LLOCUS, a novel RSS-based learning system that uses crowdsourced RF sensing to estimate the location and power of mobile transmitters in real-time while allowing mobility of the crowdsourcing nodes. The intelligence of LLOCUS comes from three different components: a novel RSS interpolation method, a regularized version of radial basis interpolation (RRBI), and a novel regression-based method for transmit power estimation. We collected extensive traces of RF data in diverse settings by creating real crowdsourcing scenarios. Our evaluations using real-world data show that LLOCUS reduces the localization error by 17-68% compared to several non-learning methods. We created an end-to-end prototype implementation of LLOCUS using Sitara SDRs.

Millimeter-wave Networks

  • mmWave Spectrum Sharing using Carrier Sensing (IEEE TWC '22): Uncoordinated spectrum sharing brings in the significant challenge of distributed interference management. We developed and analyzed a fully distributed approach for unlicensed mmWave spectrum sharing among cellular operators with no inter-operator coordination. We proposed using carrier sensing (CS) at receiver (CSR) to overcome the limitations of CS at transmitter. Using stochastic geometry, we developed a general framework for downlink coverage probability analysis of a mmWave network in the presence of CS and derived the downlink coverage probability expressions for my CS protocols. Simulations and numerical examples based on our analysis showed that our proposed enhancements led to an improvement in downlink coverage probability, compared to the case with no CS.

  • Game based mmWave Spectrum Sharing (IEEE TWC '21): We considered a similar setup as above but assumed that the time slots of all BSs are synchronized. We proposed a novel problem formulation based on the Lyapunov stochastic optimization framework and decomposed the network utility optimization into two sub-optimization problems. One of the sub-problems involves power allocation for the UEs and is stochastic and non-convex. We used a non-cooperative game-based approach to solve it in a distributed manner. Using numerical evaluations, we showed that our proposed approach improves the network utility by 15% over other MAC protocols.

  • Reinforcement Learning based mmWave Spectrum Sharing (IEEE DySpan '22): As an alternative to the game-based approach, we developed a reinforcement learning (RL), specifically Q-learning, based approach. Using simulations, we demonstrate that our RL approach achieves higher network-level utility than the game-based approach.

  • mmWave User Association and Beam Scheduling (ACM mmNets '22, submitted to IEEE TMC): We have developed an optimization framework for joint user association and beam scheduling in dense mmWave networks, with an emphasis on satisfying users' rate requirements. The framework includes an NP-hard combinatorial optimization problem, and we proposed a low-complexity sub-optimal algorithm to solve it. Unlike the above works (in the category of 'Millimeter-wave Networks') that use a distributed approach for interference management, this work uses a centralized approach.

Spatial Prediction of RF Signal

  • Proactive Spatial Prediction of Radio Environment (IEEE SECON '23): Worked on a ML-based method for proactive spatial signal strength prediction in shared spectrum wireless networks. Here ‘proactive’ implies that the spatial predictions must be done for a transmitter for which no measurement has been collected. Our proposed approach is a supervised deep learning-based method that relies on crowdsourced measurements for data collection. Our proposed method performs reasonably well in terms of signal strength prediction, ≈ 5 dB mean absolute error. Importantly, using our approach, secondary transmitters can be activated with ≈ 97% probability of not causing interference.

  • Agile Radio Map Prediction Using Deep Learning (IEEE ICASSP '23, submitted to IEEE OJ-SP): We participated in the 2023 IEEE ICASSP Signal Processing Grand Challenge, namely, the First Pathloss Radio Map Prediction Challenge. We secured the second rank in this competition.We introduced a runtime-efficient RF map prediction method based on u-net. By applying specialized and accelerated preprocessing and a customized loss function, we were able to achieves an average normalized root-mean-square error of 0.045 with an average of 14 milliseconds (ms) runtime.

Device Identification and Anomaly detection

  • Receiver Agnostic Transmitter Fingerprinting (submitted to IEEE TCCN): Was involved in developing a fingerprinting method that can be used to identify different transmitters based on the PHY layer attributes of transmitted signals. We use a GAN-based deep-learning method for robustness with respect to the fingerprinting uncertainties introduced by the receivers. Our approach improves the transmitter classification accuracy by 20% compared to the case where receivers' fingerprints are not considered.

  • Fingerprinting IoT Devices Using Latent Physical Side-Channels (ACM IMWUT '23): Was involved in development of an emanations-based fingerprinting system that can authenticate IoT devices at range. The advantage of our developed system is that we can authenticate low-power IoT devices using features intrinsic to their normal operation. Specifically, we utilize electromagnetic emanations derived from the processor’s clock to fingerprint. Our experiments demonstrate that we achieve ≥ 95% accuracy on average, applicability in a wide range of IoT scenarios, as well as support for IoT applications such as finding hidden devices.

  • Anomaly Detection in Aviation Bands (ACM TOSN '23): There is an urgency for airport facilities to acquire the capability to continuously monitor aviation bands for real-time detection of anomalies. To meet this critical need, we designed and built AviSense, an SDR-based real-time, versatile system for monitoring aviation bands. AviSense uses a novel combination of signal detection, autoencoder (AE) based unsupervised anomaly detection, and spectrum characterization to detect and distinguish anomalies. We evaluated AviSense with real-world aviation signal measurements and found that our signal classification capability achieves a true positive rate of 99% and a false positive rate of < 4%.

bottom of page