22 November 2024 in IEEE Transactions on Consumer Electronics
Along with the explosive demand for massive data computation, federated local-edge-cloud computing enables many IoT task offloading processes and has recently gained widespread attention in consumer-centric healthcare. However, existing approaches for latency-critical task offloading do not pay sufficient attention to patient privacy data and thus are far from being ready for large-scale practical deployment. In this paper, a novel local-edge-cloud based federated learning framework with one-way knowledge distillation for latency-critical and security-aware task offloading is proposed, while a task offloading scheme based on deep reinforcement learning is presented. Then, based on the consideration of privacy guarantees, computing resources and offloading decisions, an optimization problem is formulated to minimize the overall latency of all tasks for local medical devices. Additionally, a pseudo-task mechanism is adopted in the federated framework for the uploaded weights to aggregate the models on the edges and the cloud. Besides, a one-way knowledge distillation mechanism is proposed to transfer knowledge from a teacher model to student models, which compresses the network parameters and reduces the model latency. Finally, experiments based on Chest X-Ray Images of Pneumonia are conducted for corroborating the theoretical results.
18 February 2025 in IEEE Internet of Things Journal
Federated Learning (FL) is an emerging distributed machine learning paradigm that has privacy-preserving properties, but still poses privacy leakage risks in traditional centralized FL (CFL) caused by the frequent transmission of model parameters during training. Due to the resource differences among nodes, the training process is constrained by the slowest node, while frequent data transmissions result in significant communication overhead, leading to the reduced overall efficiency. What’s more, resource-rich nodes cannot fully utilize their potentials, and large models cannot be deployed on resource-constrained end devices. Therefore, selecting participating nodes and their local training strategies efficiently is a key issue in FL. To address the aforementioned issues, this paper proposes an edge-cooperative decentralized lightweight FL framework (Dec-LWFL), which introduces a multi-agent reinforcement learning (MARL) method, and designs a scoring voting mechanism for selecting participating FL nodes, determining their local training strategies, and allocating the model aggregation tasks. During agent interactions, Gaussian noise is added to the state information to safeguard the privacy of edge nodes and users, and Rényi differential privacy (RDP) is utilized to quantify the effectiveness of the privacy protection mechanism. To adapt to the resource-constrained IoT environment, Huffman coding is employed during FL training phase to compress the transmitted models and reduce the communication overhead; then, knowledge distillation is selected during the deployment phase to achieve the goal of lightweight deployment. Experimental results demonstrate that Dec-LWFL can effectively balance the privacy and performance. The framework can significantly optimize the training latency and energy consumption, while satisfying the requirements for lightweight deployment.
23 July 2025 in IEEE Transactions on Mobile Computing
With the development of Federated Learning (FL) in IoT Edge-Cloud Computing environments, mobile terminals are able to cooperate without the leakage on raw data. However, factors including the terminals’ high mobility and the network fluctuations make the cooperator selection during FL training extremely complex. Under the distributed cooperation, traditional FL strategies show certain limitations and cannot always select the available nodes when training, leading to the difficulties in energy and latency optimization. In this paper, we propose a Hierarchical Reinforcement Learning (HRL)-based federated knowledge distillation (HRL-FedKD) framework in which both high-level and low-level controllers utilize the Double Deep Q-Network (DDQN) algorithm. The high-level controller selects the nodes participating in FL training, while the low-level controller determines the number of local training epochs for each node. After training, the global model will be compressed into a lightweight model by knowledge distillation (KD) in deployment while preserving the personalization of local models. The experiments were conducted using Chest X-Ray and Brain Tumor MRI datasets to validate the proposed FL strategy. The results demonstrate that the HRL-FedKD framework can effectively optimize latency and energy consumption in complex state spaces.
1 January 2026 in IEEE Transactions on Industrial Informatics