Adaptive Privacy-Preserving Federated Learning Framework for Resource-Constrained Edge Intelligence

The rapid proliferation of edge intelligence applications has increased the demand for distributed learning frameworks capable of operating efficiently under limited computational resources while preserving user privacy. Federated Learning (FL) has emerged as a promising paradigm that enables collaborative model training without transferring raw data to centralized servers. However, conventional FL approaches often suffer from high communication overhead, limited adaptability to heterogeneous edge devices, and insufficient consideration of model efficiency and privacy requirements. This paper presents an Adaptive Privacy-Preserving Federated Learning (APPFL) framework for resource-constrained edge intelligence environments. The proposed framework jointly integrates differential privacy, model pruning, and quantization to improve learning efficiency while maintaining data confidentiality. Unlike traditional FL methods, the proposed approach dynamically adapts compression and privacy mechanisms according to device capabilities, enabling heterogeneous edge devices to participate effectively in the federated training process. Model pruning reduces computational complexity by eliminating redundant parameters, while quantization decreases memory usage and communication costs. Differential privacy is employed to protect sensitive local information during model updates. Experimental evaluations conducted on the CIFAR-۱۰ dataset under non-IID federated settings demonstrate that the proposed framework significantly reduces model size and communication overhead while maintaining competitive classification performance. The results indicate that adaptive integration of privacy preservation and model compression provides an effective balance between learning accuracy, resource efficiency, and scalability. Therefore, the proposed framework represents a practical solution for deploying intelligent services in real-world resource-constrained edge environments.

کلیدواژه ها:

Federated Learning ، Edge Intelligence ، Differential Privacy ، Model Compression ، Pruning ، Quantization ، Resource-Constrained Computing