Bahar Farahani
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Bahar Farahani has a total of 59 co-authors. Here are the top co-authors by number of shared publications:
- Farshad Firouzi
Arizona State University
6 shared publications - Fereidoon Shams
Shahid Beheshti University
3 shared publications - JaeSeung Song
Sejong University, Seoul, South Korea, Republic of Korea
3 shared publications - Krishnendu Chakrabarty
Arizona State University
3 shared publications - Reza Fotohi
Faculty of Computer Science and Engineering, Shahid Beheshti University, Tehran, Iran
3 shared publications Mahmoud Daneshmand
Stevens Institute of Technology
2 shared publications
Latest Publications
- Open Access
- 24 March 2026
- Published by Elsevier in Digital Communications and Networks
- Vol. 12 (6) , 869-893
Generative IoT: advancing IoT with generative AI and large language models
- 24 March 2026
- Published by Elsevier in Digital Communications and Networks
- Vol. 12 (6) , 869-893
Open AccessGenerative IoT: advancing IoT with generative AI and large language models
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
DiabLLM: An LLM-Based Framework for Blood Glucose Prediction in Type 1 Diabetes
Accurate Blood Glucose (BG) prediction is essential for enabling glycemic control in individuals with Type 1 Diabetes Mellitus (T1DM), particularly within Smart and Connected Health (SCH) systems that integrate Continuous Glucose Monitoring (CGM) and automated insulin delivery. The adaptability of Large Language Models (LLMs) provides a promising foundation for unified, fine-tunable forecasting models. We introduce DiabLLM, a framework based on two recent LLM-based architectures: Time-LLM, which incorporates a lightweight projection layer and alignment techniques to transform time-series data into embeddings interpretable by pre-trained LLMs, and Chronos, which employs time-series-aware tokenization and quantization to convert continuous inputs into discrete sequences for forecasting. Both models process 30-minute sequences of six historical BG values and predict 30- and 45-minute horizons. Experimental results on the OhioT1DM and D1NAMO datasets demonstrate that DiabLLM outper forms state-of-the-art baselines, including a Deep Reinforcement Learning model and an ensemble of LSTM, GRU, and WaveNet, achieving up to 27% improvement in RMSE and 37% in MAE. To enhance robustness to noisy and missing input data, a denoising autoencoder was employed for input reconstruction, yielding improved predictive performance. In addition, knowledge distillation was shown to significantly compress the model, making it a practical candidate for efficient deployment on resource-constrained edge devices without compromising accuracy.
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP (21682194) , 1-14
DiabLLM: An LLM-Based Framework for Blood Glucose Prediction in Type 1 Diabetes
Accurate Blood Glucose (BG) prediction is essential for enabling glycemic control in individuals with Type 1 Diabetes Mellitus (T1DM), particularly within Smart and Connected Health (SCH) systems that integrate Continuous Glucose Monitoring (CGM) and automated insulin delivery. The adaptability of Large Language Models (LLMs) provides a promising foundation for unified, fine-tunable forecasting models. We introduce DiabLLM, a framework based on two recent LLM-based architectures: Time-LLM, which incorporates a lightweight projection layer and alignment techniques to transform time-series data into embeddings interpretable by pre-trained LLMs, and Chronos, which employs time-series-aware tokenization and quantization to convert continuous inputs into discrete sequences for forecasting. Both models process 30-minute sequences of six historical BG values and predict 30- and 45-minute horizons. Experimental results on the OhioT1DM and D1NAMO datasets demonstrate that DiabLLM outper forms state-of-the-art baselines, including a Deep Reinforcement Learning model and an ensemble of LSTM, GRU, and WaveNet, achieving up to 27% improvement in RMSE and 37% in MAE. To enhance robustness to noisy and missing input data, a denoising autoencoder was employed for input reconstruction, yielding improved predictive performance. In addition, knowledge distillation was shown to significantly compress the model, making it a practical candidate for efficient deployment on resource-constrained edge devices without compromising accuracy.
- 26 December 2025
- Published by Springer Nature
- p. 69-90
Pandemic Resilience: Smart and Connected Solutions in Healthcare
- 26 December 2025
- Published by Springer Nature
- p. 69-90
Pandemic Resilience: Smart and Connected Solutions in Healthcare
- 7 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30 (3) , 1939-1946
Sync-GWO: Highly Private and Bandwidth-Efficient Federated Learning With a Case Study in Healthcare
Federated Learning (FL) is a transformative paradigm in machine learning that enables collaborative model training across decentralized devices, ensuring data remains securely stored locally to enhance privacy. Traditional FL techniques, such as FedAvg, rely on gradient aggregation to construct a global model. However, these approaches often incur significant communication overhead and pose privacy risks, as gradient updates can inadvertently expose sensitive information. To address these challenges, this paper presents a novel FL framework that formulates federated optimization as a Multi-Objective Optimization (MOO) problem and proposes Sync-GWO, an innovative adaptation of the Grey Wolf Optimizer (GWO) tailored specifically for FL. In contrast to conventional population-based methods that require full population transfers, Sync-GWO leverages synchronized Pseudo-Random Number Generators (PRNGs) to eliminate population data transmission. The proposed approach significantly reduces communication overhead while enhancing privacy by avoiding gradient exchange. Experimental results on COVID-19 pandemic-related Internet of Medical Things (IoMT) data demonstrate that Sync-GWO achieves up to 15% higher accuracy and 2.5× improved F1-scores compared to FedAvg, while reducing communication costs by over 99% to approximately 1KB per round. Sync-GWO is particularly well-suited for scenarios requiring high privacy and extreme communication efficiency, such as those encountered in IoMT. Furthermore, Sync-GWO exhibits robust performance on imbalanced datasets and in optimization settings involving non-differentiable objectives, where gradient-based methods are often less effective.
- 7 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30 (3) , 1939-1946
Sync-GWO: Highly Private and Bandwidth-Efficient Federated Learning With a Case Study in Healthcare
Federated Learning (FL) is a transformative paradigm in machine learning that enables collaborative model training across decentralized devices, ensuring data remains securely stored locally to enhance privacy. Traditional FL techniques, such as FedAvg, rely on gradient aggregation to construct a global model. However, these approaches often incur significant communication overhead and pose privacy risks, as gradient updates can inadvertently expose sensitive information. To address these challenges, this paper presents a novel FL framework that formulates federated optimization as a Multi-Objective Optimization (MOO) problem and proposes Sync-GWO, an innovative adaptation of the Grey Wolf Optimizer (GWO) tailored specifically for FL. In contrast to conventional population-based methods that require full population transfers, Sync-GWO leverages synchronized Pseudo-Random Number Generators (PRNGs) to eliminate population data transmission. The proposed approach significantly reduces communication overhead while enhancing privacy by avoiding gradient exchange. Experimental results on COVID-19 pandemic-related Internet of Medical Things (IoMT) data demonstrate that Sync-GWO achieves up to 15% higher accuracy and 2.5× improved F1-scores compared to FedAvg, while reducing communication costs by over 99% to approximately 1KB per round. Sync-GWO is particularly well-suited for scenarios requiring high privacy and extreme communication efficiency, such as those encountered in IoMT. Furthermore, Sync-GWO exhibits robust performance on imbalanced datasets and in optimization settings involving non-differentiable objectives, where gradient-based methods are often less effective.
- 1 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11 (19) , 31988-31998
A Lightweight and Secure Deep Learning Model for Privacy-Preserving Federated Learning in Intelligent Enterprises
The ever-growing Internet of Things (IoT) connections drive a new type of organization, the Intelligent Enterprise. In intelligent enterprises, machine learning-based models are adopted to extract insights from data. Due to these traditional models’ efficiency and privacy challenges, a new federated learning (FL) paradigm has emerged. In FL, multiple enterprises can jointly train a model to update a final model. However, firstly, FL-trained models usually perform worse than centralized models, especially when enterprises’ training data is non-IID (Independent and Identically Distributed). Second, due to the centrality of FL and the untrustworthiness of local enterprises, traditional FL solutions are vulnerable to poisoning and inference attacks and violate privacy. Thirdly, the continuous transfer of parameters between enterprises and servers increases communication costs. Therefore, to this end, the FedAnil+ model is proposed, a novel, lightweight, and secure Federated Deep LeArning Model that includes three main phases. In the first phase, the goal is to solve the data type distribution skew challenge. Addressing privacy concerns against poisoning and inference attacks is given in the second phase. Finally, to alleviate the communication overhead, a novel compression approach is proposed that significantly reduces the size of the updates. The experiment results validate that FedAnil+ is secure against inference and poisoning attacks with better accuracy. In addition, in terms of model accuracy (13%, 16%, and 26%), communication cost (17%, 21%, and 25%), and computation cost (7%, 9%, and 11%) improvements over existing approaches. The FedAnil+ code is available on GitHub.
- 1 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11 (19) , 31988-31998
A Lightweight and Secure Deep Learning Model for Privacy-Preserving Federated Learning in Intelligent Enterprises
The ever-growing Internet of Things (IoT) connections drive a new type of organization, the Intelligent Enterprise. In intelligent enterprises, machine learning-based models are adopted to extract insights from data. Due to these traditional models’ efficiency and privacy challenges, a new federated learning (FL) paradigm has emerged. In FL, multiple enterprises can jointly train a model to update a final model. However, firstly, FL-trained models usually perform worse than centralized models, especially when enterprises’ training data is non-IID (Independent and Identically Distributed). Second, due to the centrality of FL and the untrustworthiness of local enterprises, traditional FL solutions are vulnerable to poisoning and inference attacks and violate privacy. Thirdly, the continuous transfer of parameters between enterprises and servers increases communication costs. Therefore, to this end, the FedAnil+ model is proposed, a novel, lightweight, and secure Federated Deep LeArning Model that includes three main phases. In the first phase, the goal is to solve the data type distribution skew challenge. Addressing privacy concerns against poisoning and inference attacks is given in the second phase. Finally, to alleviate the communication overhead, a novel compression approach is proposed that significantly reduces the size of the updates. The experiment results validate that FedAnil+ is secure against inference and poisoning attacks with better accuracy. In addition, in terms of model accuracy (13%, 16%, and 26%), communication cost (17%, 21%, and 25%), and computation cost (7%, 9%, and 11%) improvements over existing approaches. The FedAnil+ code is available on GitHub.
- 3 June 2024
- Published by Springer Nature in Peer-to-Peer Networking and Applications
- Vol. 17 (5) , 2781-2795
Feature fusion federated learning for privacy-aware indoor localization
- 3 June 2024
- Published by Springer Nature in Peer-to-Peer Networking and Applications
- Vol. 17 (5) , 2781-2795
Feature fusion federated learning for privacy-aware indoor localization
- 7 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 10 (21) , 18553-18562
Toward a Personalized Clustered Federated Learning: A Speech Recognition Case Study
Most speech recognition systems utilize cloud computing for model training and updates. Speech data, being Personally Identifiable Information (PII), encompasses personal, privacysensitive, and regulated content. Relying on centralized servers or third parties can threaten confidential data, resulting in privacy breaches. Therefore, privacy issues and strict regulations (e.g, EU’s GDPR, California’s CCPA, and the Privacy Act in Australia) limit the availability of large datasets. The scarcity of datasets is particularly pronounced in less-represented languages, like Persian, adversely impacting innovation and data-driven product development. To overcome the challenges posed by the scarcity of datasets and privacy concerns, for the first time, we propose a novel Federated Learning (FL) solution for Persian Spoken Isolated Digit Recognition. This proposed technique bridges the gap between privacy and utility by enabling the training of an algorithm using decentralized datasets stored on edge devices or servers, without the need for data exchange. None Independent and Identically Distributed Data (non-IID), such as unique speaker accents, poses a challenge in speech recognition, especially in a FL setup. Regrettably, this challenge has largely been overlooked in existing techniques and methodologies. To address this, we present an innovative Personalized Clustered Federated Learning (PCFL) approach that successfully exploits similarities among the private data distributions and captures distinctive characteristics inherent in each client’s data in order to train models. The experimental results show that while the proposed solution significantly addresses privacy concerns, it has a negligible performance loss compared to centralized model training techniques.
- 7 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 10 (21) , 18553-18562
Toward a Personalized Clustered Federated Learning: A Speech Recognition Case Study
Most speech recognition systems utilize cloud computing for model training and updates. Speech data, being Personally Identifiable Information (PII), encompasses personal, privacysensitive, and regulated content. Relying on centralized servers or third parties can threaten confidential data, resulting in privacy breaches. Therefore, privacy issues and strict regulations (e.g, EU’s GDPR, California’s CCPA, and the Privacy Act in Australia) limit the availability of large datasets. The scarcity of datasets is particularly pronounced in less-represented languages, like Persian, adversely impacting innovation and data-driven product development. To overcome the challenges posed by the scarcity of datasets and privacy concerns, for the first time, we propose a novel Federated Learning (FL) solution for Persian Spoken Isolated Digit Recognition. This proposed technique bridges the gap between privacy and utility by enabling the training of an algorithm using decentralized datasets stored on edge devices or servers, without the need for data exchange. None Independent and Identically Distributed Data (non-IID), such as unique speaker accents, poses a challenge in speech recognition, especially in a FL setup. Regrettably, this challenge has largely been overlooked in existing techniques and methodologies. To address this, we present an innovative Personalized Clustered Federated Learning (PCFL) approach that successfully exploits similarities among the private data distributions and captures distinctive characteristics inherent in each client’s data in order to train models. The experimental results show that while the proposed solution significantly addresses privacy concerns, it has a negligible performance loss compared to centralized model training techniques.
- Open Access
- 1 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-7
Auto-tuning HyperParameters of SGD Matrix Factorization-Based Recommender Systems Using Genetic Algorithm
Recommender systems enable companies to generate meaningful recommendations to users for items or products that might interest them. Stochastic Gradient Descent Matrix Factorization (SGD-MF) is one of the most popular model-based recommender systems. Fractional Adaptive Stochastic Gradient Descent matrix factorization (FASGD-MF) is a subset of SGD-MF-based models that apply fractional calculus in an adaptive way. There are some hyperparameters in these models that impact the quality of the recommender system. However, searching the hyperparameter space to find the best configuration using an exhaustive search is often a time-consuming task. This paper employs a genetic algorithm as a search metaheuristic to tackle this problem. The proposed method is designed based on non-uniform mutation and whole arithmetic crossover. The results indicate that optimizing hyperparameters by the proposed method not only adjusts the values of hyperparameters automatically but also can improve the quality of SGD-MF-based models. Implementing the proposed genetic algorithm on two datasets (MovieLens 100K and MovieLens 1M) verifies the assertion about the performance.
- 1 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-7
Open AccessAuto-tuning HyperParameters of SGD Matrix Factorization-Based Recommender Systems Using Genetic Algorithm
Recommender systems enable companies to generate meaningful recommendations to users for items or products that might interest them. Stochastic Gradient Descent Matrix Factorization (SGD-MF) is one of the most popular model-based recommender systems. Fractional Adaptive Stochastic Gradient Descent matrix factorization (FASGD-MF) is a subset of SGD-MF-based models that apply fractional calculus in an adaptive way. There are some hyperparameters in these models that impact the quality of the recommender system. However, searching the hyperparameter space to find the best configuration using an exhaustive search is often a time-consuming task. This paper employs a genetic algorithm as a search metaheuristic to tackle this problem. The proposed method is designed based on non-uniform mutation and whole arithmetic crossover. The results indicate that optimizing hyperparameters by the proposed method not only adjusts the values of hyperparameters automatically but also can improve the quality of SGD-MF-based models. Implementing the proposed genetic algorithm on two datasets (MovieLens 100K and MovieLens 1M) verifies the assertion about the performance.