Knight Foundation School of Computing and Information Sciences
Ahmed Imteaj is currently a PhD Candidate and Graduate Assistant at the Knight Foundation School of Computing and Information Sciences, Florida International University, under the supervision of Professor M. Hadi Amini. He is also a research lab member of Sustainability, Optimization, and Learning for InterDependent networks laboratory (solid lab) at Florida International University. His research interests span federated learning, internet of things (IoT), smart systems, and Blockchain. He holds a B.Sc. degree in Computer Science and Engineering from Chittagong University of Engineering and Technology (CUET), Bangladesh in 2015. From 2015 to 2018, he worked as a Lecturer at International Islamic University Chittagong (IIUC), Chittagong, Bangladesh. Ahmed’s work on federated learning for IoT environments is the recipient of the best paper award from “2019 IEEE Conference on Computational Science & Computational Intelligence” and won the second place at 2021 Florida International University Graduate Student Appreciation Week. Ahmed has published several conference and journal papers.
The ubiquitous nature of IoT devices causes huge data streams due to their widespread applications. Storing and processing such vast amounts of data in a centralized location is costly and time-consuming. In order attain a better ML model under the conventional centralized approach, the users may need to compromise their privacy by sharing the private data with data centers. Federated Learning (FL) is emerging due to its capabilities as a distributed machine learning training over a network of available devices. Though FL has a unique way of generating a cumulative global model by learning from the client’s model parameters, it has distinctive challenges from conventional distributed optimization, such as system heterogeneity and statistical heterogeneity, particularly while applying FL in a resource-constrained IoT environments. The FL process enables network clients to generate a joint model by leveraging on-device model training of the clients using their local resources, pushing clients’ models at the edge, and learning from the global model. Due to such heterogeneity, if we deploy the FL algorithms in an IoT environment (with potentially several IoT agents that have limited local computing resources), then most of the IoT devices may turn into stragglers, and consequently, the FL process may hamper. Besides, as the IoT devices are vulnerable and prone to attacks, they may generate inappropriate models that may prolong the model convergence and exhaust the resources of the participated FL devices. In order to address these challenges, this research is focusing on how to leverage an effective FL model that can handle heterogeneous IoT devices tackling the challenges arises from systems heterogeneity and statistical heterogeneity, such as straggler issues, large data across the networks, diverge model infusion, etc. To the end, we focus on how we can apply our proposed FL algorithm in solving real-life problems such as improving resilience of resource-constrained critical infrastructures.