Sudhir Kumar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Sudhir Kumar has a total of 104 co-authors. Here are the top co-authors by number of shared publications:
- Ankur Pandey
Department of Electrical and Electronics Engineering, Rajiv Gandhi Institute of Petroleum Technology, Amethi, India
17 shared publications Rajesh M. Hegde
15 shared publications- Pritam Khan
SRM Institute of Science and Technology
8 shared publications - Piyush Tiwary
Indian Institute of Science, Bengaluru, India
6 shared publications - Ajit Kumar
Department of Electrical Engineering, Indian Institute of Technology Patna, Patna, India
5 shared publications - Gaurav Prasad
Department of Electrical Engineering, Indian Institute of Technology Patna, Patna, India
5 shared publications
Latest Publications
- 1 April 2026
- Published by Springer Nature
- p. 488-502
Wireless Device Localization in Rician Faded IRS Assisted 6G Networks
- 1 April 2026
- Published by Springer Nature
- p. 488-502
Wireless Device Localization in Rician Faded IRS Assisted 6G Networks
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 10, 35-46
Edge-Assisted Domain Adaptive Passive RF-Based Sensing for Localization and Activity Recognition
Ultra-High Frequency (UHF) Radio Frequency Identification (RFID) technology enables scalable, passive device-free location and activity recognition (LAR) in smart warehouses, assisted living environments, and industrial Internet of Things (IoT) through the deployment of commercial off-the-shelf (COTS) readers. However, it is observed that performance degrades across different environments due to domain shifts resulting from changes in shelving layouts, inventory density, metallic interference, tag orientation, worker dynamics, and multi-reader configurations. This work proposes CycleSiamese, a lightweight, class-aware domain adaptation framework suitable for passive RFID systems, which integrates cycle-consistent Received Signal Strength (RSS) transformations with a Siamese classifier, for LAR using minimal labeled target samples. Evaluated on the COTS UHF RFID floor dataset featuring Impinj Monza 4 Quiet Tag (QT) passive tags beneath an apartment floor interrogated by an Impinj Speedway Revolution R420 reader for device-free ambient assisted living monitoring, the proposed method reduces linear inter-personal domain discrepancy from 21.58 to 2.16 (88 % reduction), achieving 82.2 % Human Activity Recognition (HAR) accuracy with limited target samples. Further, the proposed method is successfully validated on three additional Radio Frequency (RF)-based datasets, which include temporal drift, device heterogeneity, and dynamic environments, while supporting real-time edge deployment on resource-constrained RFID gateways.
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 10, 35-46
Edge-Assisted Domain Adaptive Passive RF-Based Sensing for Localization and Activity Recognition
Ultra-High Frequency (UHF) Radio Frequency Identification (RFID) technology enables scalable, passive device-free location and activity recognition (LAR) in smart warehouses, assisted living environments, and industrial Internet of Things (IoT) through the deployment of commercial off-the-shelf (COTS) readers. However, it is observed that performance degrades across different environments due to domain shifts resulting from changes in shelving layouts, inventory density, metallic interference, tag orientation, worker dynamics, and multi-reader configurations. This work proposes CycleSiamese, a lightweight, class-aware domain adaptation framework suitable for passive RFID systems, which integrates cycle-consistent Received Signal Strength (RSS) transformations with a Siamese classifier, for LAR using minimal labeled target samples. Evaluated on the COTS UHF RFID floor dataset featuring Impinj Monza 4 Quiet Tag (QT) passive tags beneath an apartment floor interrogated by an Impinj Speedway Revolution R420 reader for device-free ambient assisted living monitoring, the proposed method reduces linear inter-personal domain discrepancy from 21.58 to 2.16 (88 % reduction), achieving 82.2 % Human Activity Recognition (HAR) accuracy with limited target samples. Further, the proposed method is successfully validated on three additional Radio Frequency (RF)-based datasets, which include temporal drift, device heterogeneity, and dynamic environments, while supporting real-time edge deployment on resource-constrained RFID gateways.
- 17 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 9, 797-806
Deep Learning-Based Secure Tag Selection in BackCom Network With RIS-Induced Interference
This article investigates the secrecy performance of a non-linear energy-harvesting backscatter communication (BackCom) network in the presence of direct link and reconfigurable intelligent surface (RIS) interference. The network comprises a source, multiple passive tags, an RIS, and a legitimate reader, with an eavesdropper attempting to intercept the communication. We analyze a tag selection scheme based on long-short-term memory (LSTM) to address the challenge of selecting tags under the influence of direct link and the RIS interference. The nonideal behavior of the RIS is exploited to enhance secrecy performance by modeling RIS phase errors using Von Mises and uniform distributions. Because of interference from the direct link and the RIS being common to all tags, the secrecy rates of different tags are correlated. The LSTM-based scheme effectively captures this correlation and perfectly matches the conventional selection scheme on low and high tag counts. The secrecy outage probability (SOP) achieved using the LSTM outperforms other machine learning techniques, such as k-nearest neighbors (k-NN), decision trees (DT), and support vector machines (SVM). We also demonstrate the impact of RIS elements, phase error parameters, and the number of tags on the SOP in the considered RIS-aided BackCom network.
- 17 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 9, 797-806
Deep Learning-Based Secure Tag Selection in BackCom Network With RIS-Induced Interference
This article investigates the secrecy performance of a non-linear energy-harvesting backscatter communication (BackCom) network in the presence of direct link and reconfigurable intelligent surface (RIS) interference. The network comprises a source, multiple passive tags, an RIS, and a legitimate reader, with an eavesdropper attempting to intercept the communication. We analyze a tag selection scheme based on long-short-term memory (LSTM) to address the challenge of selecting tags under the influence of direct link and the RIS interference. The nonideal behavior of the RIS is exploited to enhance secrecy performance by modeling RIS phase errors using Von Mises and uniform distributions. Because of interference from the direct link and the RIS being common to all tags, the secrecy rates of different tags are correlated. The LSTM-based scheme effectively captures this correlation and perfectly matches the conventional selection scheme on low and high tag counts. The secrecy outage probability (SOP) achieved using the LSTM outperforms other machine learning techniques, such as k-nearest neighbors (k-NN), decision trees (DT), and support vector machines (SVM). We also demonstrate the impact of RIS elements, phase error parameters, and the number of tags on the SOP in the considered RIS-aided BackCom network.
- 29 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Network
- Vol. 40 (1) , 140-145
Knowledge Distillation-Based AIoT Framework for Efficient Wireless Gesture Sensing in B5G/6G Networks
The advancements in 5G and 6G networks have opened new possibilities for wireless sensing in smart environments. Gesture and activity recognition are crucial for interactive systems, particularly in Internet of Things (IoT) applications, where real-time performance and cost efficiency are critical. This article presents a novel Knowledge Distillation (KD)-based Artificial Intelligence of Things (AIoT) framework that enables efficient, lightweight models for gesture and activity recognition using Wireless Fidelity (Wi-Fi) Channel State Information (CSI). Unlike conventional camera-based or wearable sensor methods, Wi-Fi-based sensing offers a low-cost, privacy-preserving alternative suitable for security-sensitive applications in B5G/6G networks. Experimental results demonstrate that despite its compact structure, the proposed KD-based model achieves accuracy comparable to state-of-the-art methods while being optimized for resource-constrained edge devices. These findings underscore the potential of KD to develop robust, scalable, and energy-efficient AIoT solutions for next-generation wireless sensing.
- 29 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Network
- Vol. 40 (1) , 140-145
Knowledge Distillation-Based AIoT Framework for Efficient Wireless Gesture Sensing in B5G/6G Networks
The advancements in 5G and 6G networks have opened new possibilities for wireless sensing in smart environments. Gesture and activity recognition are crucial for interactive systems, particularly in Internet of Things (IoT) applications, where real-time performance and cost efficiency are critical. This article presents a novel Knowledge Distillation (KD)-based Artificial Intelligence of Things (AIoT) framework that enables efficient, lightweight models for gesture and activity recognition using Wireless Fidelity (Wi-Fi) Channel State Information (CSI). Unlike conventional camera-based or wearable sensor methods, Wi-Fi-based sensing offers a low-cost, privacy-preserving alternative suitable for security-sensitive applications in B5G/6G networks. Experimental results demonstrate that despite its compact structure, the proposed KD-based model achieves accuracy comparable to state-of-the-art methods while being optimized for resource-constrained edge devices. These findings underscore the potential of KD to develop robust, scalable, and energy-efficient AIoT solutions for next-generation wireless sensing.
- 18 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (11) , 18423-18426
Cramér-Rao Bound Analysis for Localization in α-μ Fading IoT Environments
The range-based localization approaches using Received Signal Strength (RSS) are inaccurate due to randomness in received signal power arises from shadowing and multipath effects. By modeling multipath effects using a generalized fading environment, such as α-μ distribution improves localization accuracy while maintaining low model complexity. We derive estimator’s localization error bound for model considering shadowing and multipath effects as Gaussian and α-μ distribution respectively. We also compare the location estimator performance to assess potential areas of algorithm improvement and parameters on which CRB depends. Furthermore, this paper also provides parameters that influence localization performance based on experiments performed on two real-world testbeds.
- 18 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (11) , 18423-18426
Cramér-Rao Bound Analysis for Localization in α-μ Fading IoT Environments
The range-based localization approaches using Received Signal Strength (RSS) are inaccurate due to randomness in received signal power arises from shadowing and multipath effects. By modeling multipath effects using a generalized fading environment, such as α-μ distribution improves localization accuracy while maintaining low model complexity. We derive estimator’s localization error bound for model considering shadowing and multipath effects as Gaussian and α-μ distribution respectively. We also compare the location estimator performance to assess potential areas of algorithm improvement and parameters on which CRB depends. Furthermore, this paper also provides parameters that influence localization performance based on experiments performed on two real-world testbeds.
- 27 December 2024
- Published by Springer Nature
- p. 88-101
Spectral Features-Based Machine Learning Approach to Detect SARS-COV-2 Infection Using Cough Sound
- 27 December 2024
- Published by Springer Nature
- p. 88-101
Spectral Features-Based Machine Learning Approach to Detect SARS-COV-2 Infection Using Cough Sound
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 11 (6) , 6207-6218
A Unified $\alpha \! - \! \eta \! -\! \kappa \! - \! \mu$ Fading Model Based Real-Time Localization on IoT Edge DevicesWi-Fi-based localization using Received Signal Strength (RSS) is widely adopted due to its cost-effectiveness and ubiquity. However, localization accuracy of RSS-based localization degrades due to random fluctuations from shadowing and multipath fading effects. Existing fading distributions like Rayleigh, $\kappa - \mu$ , and $\alpha$ -KMS struggle to capture all factors contributing to fading. In contrast, the $\alpha - \eta - \kappa - \mu$ distribution offers the most generalized coverage of fading in literature. However, as fading distributions become more generalized, their computational demands also increases. This results in a trade-off between localization accuracy and complexity, which is undesirable for real-time localization. In this work, we propose a novel localization strategy utilizing the $\alpha - \eta - \kappa - \mu$ distribution combined with a novel approximation method that significantly reduces computational overhead while maintaining accuracy. Our proposed strategy effectively mitigates the trade-off between localization accuracy and complexity, outperforming existing state-of-the-art (SOTA) localization techniques on simulated and real-world testbeds. The proposed strategy achieves accurate localization with a speedup of 280 times over non-approximated methods. We validate its feasibility for real-time tasks on low-compute edge device Raspberry Pi Zero W, where it demonstrates fast and accurate localization, making it suitable for real-time edge applications.
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 11 (6) , 6207-6218
A Unified $\alpha \! - \! \eta \! -\! \kappa \! - \! \mu$ Fading Model Based Real-Time Localization on IoT Edge DevicesWi-Fi-based localization using Received Signal Strength (RSS) is widely adopted due to its cost-effectiveness and ubiquity. However, localization accuracy of RSS-based localization degrades due to random fluctuations from shadowing and multipath fading effects. Existing fading distributions like Rayleigh, $\kappa - \mu$ , and $\alpha$ -KMS struggle to capture all factors contributing to fading. In contrast, the $\alpha - \eta - \kappa - \mu$ distribution offers the most generalized coverage of fading in literature. However, as fading distributions become more generalized, their computational demands also increases. This results in a trade-off between localization accuracy and complexity, which is undesirable for real-time localization. In this work, we propose a novel localization strategy utilizing the $\alpha - \eta - \kappa - \mu$ distribution combined with a novel approximation method that significantly reduces computational overhead while maintaining accuracy. Our proposed strategy effectively mitigates the trade-off between localization accuracy and complexity, outperforming existing state-of-the-art (SOTA) localization techniques on simulated and real-world testbeds. The proposed strategy achieves accurate localization with a speedup of 280 times over non-approximated methods. We validate its feasibility for real-time tasks on low-compute edge device Raspberry Pi Zero W, where it demonstrates fast and accurate localization, making it suitable for real-time edge applications.
- 23 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (1) , 1
Non-Data-Aided SNR Estimation for Molecular Communication Systems in the Internet of Bio-Nano Things
This paper proposes a non-data-aided (NDA) signal-to-noise ratio (SNR) estimation method for molecular diffusive communication channels. In contrast to previous data-aided (DA) approaches, which require a known data sequence to be shared between transmitter and receiver nanomachines, the proposed NDA method eliminates the need for prior knowledge of data symbols for the estimation process. The absence of a training sequence significantly reduces resource costs by minimizing the number of molecules required for information transmission. The expectation-maximization (EM) algorithm was used to iteratively find the maximum-likelihood (ML) estimates. The proposed method is particularly advantageous in scenarios in which accessing known data sequences is not possible, thereby ensuring robust communication in the Internet of bio-nano things. By employing the ML estimates obtained through the EM algorithm, the probabilities associated with the transmission of data symbols were derived, and the bit error probabilities were obtained to assess the system performance. The numerical results validated the effectiveness of the proposed NDA SNR estimator in accurately estimating SNR over a diverse range of conditions. The proposed NDA estimator demonstrated resilience by maintaining a low bit error probability.
- 23 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (1) , 1
Non-Data-Aided SNR Estimation for Molecular Communication Systems in the Internet of Bio-Nano Things
This paper proposes a non-data-aided (NDA) signal-to-noise ratio (SNR) estimation method for molecular diffusive communication channels. In contrast to previous data-aided (DA) approaches, which require a known data sequence to be shared between transmitter and receiver nanomachines, the proposed NDA method eliminates the need for prior knowledge of data symbols for the estimation process. The absence of a training sequence significantly reduces resource costs by minimizing the number of molecules required for information transmission. The expectation-maximization (EM) algorithm was used to iteratively find the maximum-likelihood (ML) estimates. The proposed method is particularly advantageous in scenarios in which accessing known data sequences is not possible, thereby ensuring robust communication in the Internet of bio-nano things. By employing the ML estimates obtained through the EM algorithm, the probabilities associated with the transmission of data symbols were derived, and the bit error probabilities were obtained to assess the system performance. The numerical results validated the effectiveness of the proposed NDA SNR estimator in accurately estimating SNR over a diverse range of conditions. The proposed NDA estimator demonstrated resilience by maintaining a low bit error probability.