22 December 2020 in IEEE Internet of Things Journal
In this paper, we propose a fog-enabled federated learning framework – FogFL – to facilitate distributed learning for delay-sensitive applications in resource-constrained IoT environments. While Federated Learning (FL) is a popular distributed learning approach, it suffers from communication overheads and high computational requirements. Moreover, global aggregation in FL relies on a centralized server, prone to malicious attacks, resulting in inefficient training models. We address these issues by introducing geospatially placed fog nodes into the FL framework as local aggregators. These fog nodes are responsible for defined demographics, which help share location-based information for applications with similar environments. Further, we formulate a greedy heuristic approach for selecting an optimal fog node for assuming a global aggregator’s role at each round of communication between the edge and cloud, thereby reducing the dependence on the execution at the centralized server. Fog nodes in the FogFL framework reduce communication latency and energy consumption of resource-constrained edge devices without affecting the global model’s convergence rate, thereby increasing the system’s reliability. Extensive deployment and experimental results corroborate that, in addition to a decrease in global aggregation rounds, FogFL reduces energy consumption and communication latency by 92% and 85%, respectively, as compared to state-of-the-art.
26 October 2022 in IEEE Transactions on Parallel and Distributed Systems
This work presents an efficient data-centric client selection approach, named DICE, to enable federated learning (FL) over distributed edge networks. Prior research focused on assessing the computation and communication ability of the client devices for selection in FL. On-device data quality, in terms of data volume and heterogeneity, across these distributed devices is largely overlooked. The obvious outcome is the selection of an improper subset of clients with poor-quallity data, which inevitably results in an inefficient trained model. With an aim to address this problem, in this work, we design DICE which prioritizes the data quality of the client devices in the selection phase, in addition to their computation and communication abilities, to improve the accuracy of FL. Additionally, in DICE, we introduce the assistance of vicinal edge devices to account for the lack of computation or communication abilities in certain devices without violating the privacy-preserving guarantees of FL. Towards this aim, we propose a scheme to decide the optimal edge device, in terms of latency and workload, to be selected as the helper device. The experimental results show that DICE improves convergence speed for a given level of model accuracy. Further, the simulation results show that DICE reduces delay by at least 16%, energy consumption by at least 17%, and packet loss by at least 55% compared to the existing benchmarks while prioritizing the on-device data quality across clients.
31 December 2020 in IEEE Sensors Journal
This work presents the design of an efficient edge-empowered sensor-cloud architecture equipped with a smart virtual sensing scheme for precision agriculture. Traditionally, in agricultural sensor-cloud, sensor nodes send raw sensed data periodically to the cloud, resulting in higher latency and higher energy and bandwidth consumption. The environment-dependent nature of agricultural parameters also limits the serviceability of sensor-cloud in regions with damaged or unemployed sensors. Moreover, the agricultural sensor-cloud suffers from privacy issues due to the sharing of sensitive farming data across third-party service providers. To address these drawbacks, we first propose a modified sensor-cloud architecture using edge devices as the middleware layer for sensor virtualization, thereby reducing service provisioning latency and resource consumption. Next, we propose DLSense, a novel intelligent virtualization scheme to aid in the design of virtual sensors in the absence of working sensor nodes in a region. DLSense utilizes correlation theory and distributed learning in the edge devices to predict sensor data and enables sharing of information of the trained models instead of raw sensed data, thus imparting privacy. Finally, we evaluate the performance of the DLSense scheme through extensive simulations and an experimental case study of an agricultural application. Results demonstrate that our proposed scheme reduces latency and service cost by 81% and 66%, respectively, and increases service availability by 39% compared to the state-of-the-art methods.
13 October 2021 in IEEE Transactions on Network and Service Management
NRNarendra Singh Raghuwanshi
Narendra Singh Raghuwanshi
This paper explores the concept of AI-based Communication-as-a-Service (ACUTE) to reduce transmission delay and energy consumption, while transmitting data from end-devices to the cloud in the context of 5.0. 5.0 revolutionizes connected living with the help of a unified system that provides fully automated and end-to-end services, while addressing the demands of all the citizens or users in a society. On the other hand, 6G is one of the promising communication platforms that offers the communication requirements of 5.0 by provisioning dense network deployment and fast data delivery. Building 5.0 founded on the 6G architecture enables serialized data transmission in the connected living fabric by allowing a user to connect with an access point and transmit data over a single path. Without concurrent and intelligent data transmission, the communication framework of Society 5.0 increases network delay and overall energy consumption and affects the Quality-of-Service (QoS). To address these issues, we propose a solution founded on the concept of Communication-as-a-Service (CaaS), which offers an architecture to facilitate intelligent access point virtualization for enabling concurrency in data transmission across individual users in a 6G-enabled 5.0. In ACUTE, a virtual module (VM) employed at each edge device performs concurrent data transmissions by associating with a virtual access point (VAP), which is a set of access points optimally selected using Fuzzy C-Means. Thereafter, the VM forms a virtual path (VP), which maps to a set of paths between physical access points and VAPs. ACUTE distributes the data through the VAP and associated VP and randomizes data sequence for transmission across VP. Experimental results show that ACUTE outperforms the state-of-the-art while reducing the network delay by 27%, energy consumption by 95%, packet loss by 95%, and service cost by 26%.