AWE-DPFL: Adaptive weighting and dynamic privacy budget federated learning for heterogeneous data in IoT

Abstract

In the era of data-driven artificial intelligence, the widespread deployment of IoT devices has amplified concerns around privacy and data security. Federated learning (FL) offers a promising solution by enabling local model training without exposing raw data, effectively mitigating privacy risks. However, the inherent heterogeneity of IoT data leads to significant disparities in data distributions across different clients, negatively impacting the global model’s performance. Furthermore, conventional fixed differential privacy mechanisms lack the adaptability needed to dynamically adjust to the evolving requirements of different training phases, limiting their effectiveness in privacy-preserving federated learning. To address these challenges, we propose a federated learning framework called AWE-DPFL, which integrates adaptive weight fusion and dynamic privacy budget adjustment mechanisms. AWE-DPFL employs a dynamic privacy budget adjustment strategy to allocate privacy budgets based on the variance in client model updates, thereby improving model performance while ensuring robust privacy protection. Additionally, the adaptive weight fusion mechanism assigns different weights to each client’s model, taking into account data heterogeneity and quality, which leads to an enhanced global model that better reflects individual client contributions. Moreover, AWE-DPFL incorporates meta-learning alongside differential privacy techniques during local model training, resulting in an effective balance between data privacy and model performance. This approach not only improves model adaptability and generalization across diverse data distributions but also ensures that privacy requirements are met throughout the training process. Experimental evaluations demonstrate that AWE-DPFL significantly outperforms existing approaches on the MNIST, FashionMNIST, HAR, and Edge-IIoTset datasets, showcasing its effectiveness as a federated learning solution for real-world IoT applications.