Dynamic Adaptive FedCOM: A Novel Communication-Efficient Approach for Federated Learning

Dynamic Adaptive FedCOM: A Novel Communication-Efficient Approach for Federated Learning

연구 분야: Infrastructure

학회: SN Computer Science

Federated learning presents a paradigm where multiple decentralized edge devices collaboratively train a machine learning model while keeping all training data localized, thus addressing significant concerns regarding data privacy and security. However, the efficacy of federated learning is often limited by substantial communication overhead, especially when scaling to a large number of nodes with diverse data distributions. This paper introduces the Dynamic Adaptive FedCOM algorithm, a novel approach designed to enhance communication efficiency in federated learning environments. Dynamic Adaptive FedCOM dynamically adjusts data compression and aggregation weights based on real-time network conditions and the quality of data from individual nodes. This adaptive mechanism not only reduces the required network bandwidth but also ensures that the learning process remains robust against variations in data quality and node reliability. We evaluate the performance of Dynamic Adaptive FedCOM using three diverse datasets: CIFAR-10, MNIST and the Enron Email Dataset, which represent challenges across image classification (CIFAR-10, MNIST) and text processing domains (Enron Email). The evaluation was conducted using a fixed 80/20 train-test split under a controlled Independent and Identically Distributed (IID) client setup with consistent hyperparameters. Our experimental evaluation demonstrates that Dynamic Adaptive FedCOM significantly reduces the number of communication rounds required to reach target accuracy levels compared to traditional methods. Specifically, it achieves a 1.51 —reduction in communication rounds compared to FedAvg, 1.36 —compared to FedDropout, and 1.06 —compared to CMFL. Furthermore, it enhances model accuracy, achieving 0.94 on CIFAR-10, 0.95 on MNIST, and 0.89 on the Enron Email dataset, outperforming existing federated learning techniques. These results highlight Dynamic Adaptive FedCOM’s potential to facilitate more scalable and efficient federated learning solutions, making it suitable for applications ranging from mobile computing to privacy-sensitive data analytics.