The rapid advancement of digital technologies and the widespread adoption of the Internet of Things (IoT) and smart devices have led to a continuous increase in the generation of sensitive data containing personal information. To utilize such sensitiv...
The rapid advancement of digital technologies and the widespread adoption of the Internet of Things (IoT) and smart devices have led to a continuous increase in the generation of sensitive data containing personal information. To utilize such sensitive data for artificial intelligence training, conventional centralized learning approaches inevitably require transmitting data directly to a central server, which raises issues such as privacy infringement risks and increased data management costs. To overcome these limitations, Federated Learning has emerged as a paradigm that enables the training of a global model without sharing raw data externally. However, since data in federated learning are stored locally on each client and are not shared, real-world environments typically exhibit Non-Independent and Identically Distributed (Non-IID) data distributions, which in turn cause performance degradation of the global model.
In this study, to address these challenges, we clearly define two structural limitations of the existing federated learning algorithm, FedAvg (Federated Averaging), and propose a new learning strategy to overcome them. First, the conventional approach of randomly selecting participating clients fails to consider the quality and contribution of client data, resulting in inefficiency. To address this, we design an intelligent client selection strategy based on reinforcement learning that incorporates data diversity, recent local loss, and contribution scores. Second, the aggregation method in FedAvg, which computes a simple weighted average proportional to the amount of client data, is insufficient to correct Non-IID bias. Therefore, we propose an exploratory aggregation method that applies the Particle Swarm Optimization (PSO) algorithm to search for an optimized global model. Through this approach, we aim to mitigate inter-client data bias in Non-IID environments and achieve more stable and higher accuracy.
Experimental results on three datasets—CIFAR-10, CIFAR-100, and FEMNIST—demonstrate that the proposed federated learning model based on reinforcement learning and PSO consistently outperforms the conventional FedAvg algorithm across all datasets. In particular, on the FEMNIST dataset configured for Non-IID experiments, where each client possesses fewer than five classes and the data distribution is extremely skewed, FedAvg fails to converge altogether, whereas the proposed algorithm converges stably, achieving a maximum accuracy of 68.4%. Furthermore, on CIFAR-10 and CIFAR-100, the proposed method achieves accuracy improvements of up to 24.1% compared to FedAvg. These results indicate that the two proposed enhancement strategies—reinforcement-learning-based client selection and PSO-based optimal model search aggregation—are effective in significantly improving the performance of federated learning under Non-IID conditions.