KKKuljeet Kaur

Kuljeet Kaur

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Recognitions

Top Cited Young Scholar (2023)
Top Cited Scholar (2023)

Kuljeet Kaur has a total of 124 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Federated Multimodal Bi-LSTM for Agentic Handover in 6G Heterogeneous Networks

    by Wei-Che Chien, Yu Huang, Cheng Dai, Xingang Liu, Kuljeet Kaur, Mubarak Alrashoud

  • Intelligent Semantic Agents Empowered Internet of Vehicles Towards 6G

    by Haotong Cao, Miao Liu, Kuljeet Kaur, Mohsen Guizani, Mubarak Alrashoud

  • Open Access

    A Hybrid AE–WGAN–HCBOU Framework for Rare Attack Augmentation and Class-Imbalanced Intrusion Detection

    by Nassim Youcef Essaidi, Ishfaq Bashir Sofi, Kuljeet Kaur, Anjali Agarwal

    The rapid expansion of digital infrastructures and the increasing sophistication of cyberattacks have made Intrusion Detection Systems (IDS) a critical component of modern network security. A major challenge in this domain is the extreme imbalance among attack classes in real-world datasets, where critical attack categories often constitute less than 10\% of total traffic. This imbalance leads to poor detection performance for rare attacks, as machine learning and deep learning models are biased toward majority classes. Traditional oversampling methods fail to capture the complex non-linear structure of minority distributions, while classical generative models suffer from instability and mode collapse when trained on sparse data. To address these limitations, this paper proposes a novel Hierarchical Cluster-Based Over- and Under-Sampling (HCBOU) mechanism that performs distribution-aware allocation of synthetic samples in latent space, ensuring balanced coverage across fragmented minority-class regions. Unlike existing approaches that focus solely on generating synthetic data, the proposed method explicitly controls where synthetic samples are placed, thereby preserving the intrinsic structure of the minority manifold. To support this, an autoencoder is employed to learn a compact and noise-reduced latent representation, and a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) is used to generate stable and diverse synthetic samples. HCBOU then enforces cluster-level density balancing, preventing oversampling in dense regions and improving representation in sparse regions. The proposed framework is evaluated on three benchmark intrusion detection datasets using multiple classifier backbones. Experimental results demonstrate significant improvements in recall and F1-score for rare attack classes, highlighting the effectiveness of distribution-aware augmentation over traditional oversampling and standard GAN-based methods.

  • Open Access

    ProHome: Recurrent Proximal Policy Optimization-driven Advanced Home Energy Management

    by Mahmoud Sallam, Kuljeet Kaur

    Advanced Home Energy Management Systems (HEMS) are essential for integrating Renewable Energy Sources (RES) and Electric Vehicles (EVs) into smart homes; however, many existing solutions overlook the temporal dependencies in energy dynamics and are evaluated under overly simplified or fully known price signals. This work proposes ProHome, a HEMS framework based on Recurrent Proximal Policy Optimization (RPPO), where an LSTM-enabled policy learns memory-dependent control strategies from sequential observations. ProHome coordinates a comprehensive set of household assets (HVAC, electric water heater, deferrable appliances, PV generation, an energy storage system, and an EV with bidirectional power exchange) under time-varying electricity prices without assuming access to future prices. A realistic smart home simulation environment driven by historical data was developed for training and testing. Results on a 1,089-day dataset (train/test split 75/25) show that ProHome achieves the lowest average daily electricity cost among the compared DRL baselines and improves EV readiness at departure. Compared with standard PPO, the proposed recurrent agent reduces the mean daily operating cost by 1% and increases the EV state of energy at departure by 2.1 percentage points toward the user-defined target, while maintaining near-zero comfort violations for HVAC and hot-water temperature.

  • GAN-Empowered Parasitic Covert Communication: Data Privacy in Next-Generation Networks

    by Pengcheng Guo, Zhi Lin, Haotong Cao, Yifu Sun, Kuljeet Kaur, Sherif Moussa

    The widespread integration of artificial intelligence (AI) in next-generation communication networks poses a serious threat to data privacy while achieving advanced signal processing. Eavesdroppers can use AI-based analysis to detect and reconstruct transmitted signals, leading to serious leakage of confidential information. In order to protect data privacy at the physical layer, we redefine covert communication as an active data protection mechanism. We propose a new parasitic covert communication framework in which communication signals are embedded into dynamically generated interference by generative adversarial networks (GANs). This method is implemented by our CDGUBSS (complex double generator unsupervised blind source separation) system. The system is explicitly designed to prevent unauthorized AI-based strategies from analyzing and compromising signals. For the intended recipient, the pretrained generator acts as a trusted key and can perfectly recover the original data. Extensive experiments have shown that our framework achieves powerful covert communication, and more importantly, it provides strong defense against data reconstruction attacks, ensuring excellent data privacy in next-generation wireless systems.

  • Energy Consumption Forecasting in Net-Zero Energy Buildings: Firefly-Driven LSTM for Smart Consumer Electronics and Edge-Based Energy Optimization

    by Mahmoud Sallam, Deepika, Kuljeet Kaur, Neeru Jindal, Georges Kaddoum, Mukesh Singh, Yatindra Nath Singh

    Net-Zero Energy Buildings (NZEBs) are present-day constructions capable of generating sufficient clean energy for their consumption. Since these buildings tend to rely on clean energy from Renewable Energy Sources (RESs), there is a challenge of controlling when and how much energy is generated. Thus, energy consumption prediction is essential to manage supply and demand using RESs and energy storage solutions. However, considering that the underlying electrical appliances have uncertain and non-linear energy consumption patterns, effectively predicting energy consumption is a tedious task. Accordingly, the existing prediction schemes, such as Random Forest Regressor (RFR), Support Vector Regressor (SVR), and Long-Short-Term Memory (LSTM), face various problems like vanishing gradients, limited memory cells, and fixed-length inputs. To overcome these issues, in this study, we propose a hybrid framework that combines the LSTM and Firefly (FF) optimization algorithms. LSTM learns complex long-term dependencies and efficiently predicts energy variations. In addition, the standard FF is modified (thereafter referred to as modified FF, MFF) to address existing issues, such as premature convergence and limited global search in complex time-series data. More specifically, FF is modified with additional components, such as chaotic logistic maps, adaptive inertia weight, and levy flight. These components generate an initially diverse population of fireflies, adjust attractiveness parameters, regulate local and global exploration capabilities, and accelerate local search by creating new best solutions. Due to these modifications, the proposed combination of LSTM and MFF converges faster, requires fewer iterations, and requires less processing time. For a comprehensive evaluation of the proposed work, the simulation results are analyzed using the Portuguese house time-series dataset. The results prove the efficacy of the proposed methodology as compared to the existing schemes. Optimized prediction of energy consumption can be used to synchronize energy supply and demand, as well as to facilitate green buildings to create a more sustainable environment. The results demonstrate the effectiveness of the proposed methodology, achieving an RMSE of 23.55 W and an R² of 0.99, thereby surpassing existing approaches. Moreover, the proposed method can be integrated with smart consumer electronics and edge devices, enabling real-time, energy-efficient decision-making at the edge for enhanced control in NZEBs and beyond.


  • Open Access

    Impact Analysis of Algorithm Optimization on Robot Deep Learning Localization Model for Real-Time Execution

    by Cédric Melançon, Maarouf Saad, Kuljeet Kaur, Julien Gascon-Samson

    The increased demand for autonomous robots in industries such as healthcare, manufacturing, and logistics is driven by the need to address labor shortages and enhance operational efficiency. A critical aspect of these robots is their ability to navigate complex environments autonomously, which relies heavily on multiple artificial intelligence models, including visual odometry. Visual odometry enables robots to estimate their motion by analyzing visual data, making it essential for navigation and obstacle avoidance. However, ensuring such models operate within real-time constraints is paramount for the robot’s responsiveness, safety, and overall functionality. The complexity and depth of such a model, which utilizes camera images and other sensor data, make it challenging to achieve real-time performance. Thus, this paper thoroughly evaluates the performance of a deep-learning visual odometry model, with a focus on its execution time and computational overhead. Furthermore, it proposes an optimized implementation to meet the stringent real-time requirements for safe and efficient robot operation. The optimization is achieved through algorithm improvement to preserve the model structure and accuracy. Multiple hardware platforms were utilized to demonstrate the resource differences between edge, fog, and cloud deployments, thereby validating the effect of the proposed model optimization. The resource usage is also compared to see the impact of the modification on multiple aspects, including computation, memory, temperature, and energy. The resulting model operationalization improves the execution time while highlighting limitations for some hardware configurations.

  • Fed-Reputed: Reputation-Aware Client Selection in Hierarchical Federated Learning for Consumer Electronics

    by M. A. Moyeen, Kuljeet Kaur, Anjali Agarwal, S. Ricardo Manzano, Marzia Zaman, Nishith Goel

    Federated Learning (FL) enables collaborative model training across distributed consumer devices without requiring raw data to be centralized—an essential feature for privacy-critical applications such as biomedical monitoring and mobile health diagnostics. However, real-world deployments often face challenges in client selection, especially in the presence of misbehaved clients. Straggler clients with poor connectivity or low resources delay convergence, while malicious clients pose threats to model integrity by injecting poisoned updates. These problems are amplified in dynamic, heterogeneous environments such as those involving consumer biomedical devices. To address this, we propose Fed-Reputed, a reputation-aware client selection framework tailored for Hierarchical Federated Learning (HFL). Unlike existing reputation-based approaches that suffer from herding effects, cold-start limitations, and imbalanced classification datasets, Fed-Reputed integrates a modified Bellman equation within a Deep Q-Learning framework. This formulation guides client selection using an Imbalanced Classification Markov Decision Process (ICMDP), leveraging both device capability and historical behavior. Extensive simulations using MNIST and FMNIST datasets under varying percentages of straggler and malicious clients demonstrate that Fed-Reputed achieves up to 50% higher global model accuracy and 1.7 times faster convergence compared to state-of-the-art selection methods. Moreover, it significantly improves the detection of misbehaving clients without sacrificing fairness or scalability. These results highlight the potential of Fed-Reputed to advance secure, scalable, and personalized Federated Learning in health-sensitive consumer environments.


  • Open Access

    HiSO-CoMA: Hierarchical Self-Optimizing Framework for O-RAN Slicing Using Cooperative Multiple Agent Deep Reinforcement Learning

    by Ohood Sabr, Georges Kaddoum, Kuljeet Kaur

    Network slicing (NS) is a cornerstone technology for sixth-generation (6G) networks, enabling the support of heterogeneous services with diverse quality-of-service (QoS) requirements. However, existing radio access network (RAN) slicing schemes often rely on single-level resource allocation, limiting their adaptability to the dynamic nature of RAN and the efficient use of limited radio resources. This leads to challenges in satisfying service-level agreements (SLAs). Moreover, effective hierarchical slicing that operates under fluctuating traffic loads, and hardware impairments for multiple antenna systems remains a challenge. To address these issues, we propose a hierarchical self-optimization framework aimed at maximizing both the long-term QoS and the spectral efficiency. Specifically, the proposed framework consists of two slicing management schemes: a cooperative multiple actor-critic (CoMA2C) scheme to manage the power and bandwidth among heterogeneous slices on a large scale. Concurrently, a multiagent deep Q-network (MADQN) scheme manages the power and beamforming for active users within each slice on a small time scale, accounting for hardware impairments, user mobility, traffic fluctuations, and channel variations. The DQN and A2C algorithms are employed in the design of the proposed schemes owing to their proven effectiveness in real-time decision-making in dynamic environments. Furthermore, a promising scheme based on rate-splitting multiple access (RSMA) is investigated for heterogeneous services. Simulation results showcase the effectiveness of our proposed framework, demonstrating its ability to satisfy SLAs for heterogeneous services while reducing network overhead and outperforming existing state-of-the-art approaches.