@Article{cmc.2026.085467,
AUTHOR = {Fangfang Shan, Yuhang Liu, Lulu Fan, Zhuo Chen, Yifan Mao, Peixue Wang},
TITLE = {SAM-ADPFL: A Geometry-Aware Adaptive Framework for Privacy-Preserving Federated Learning Systems},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27886},
ISSN = {1546-2226},
ABSTRACT = {The engineering of Federated Learning (FL) systems faces significant challenges in balancing two critical non-functional requirements: ensuring robust system utility and maintaining high privacy protection standards under non-independent and identically distributed (Non-IID) data environments. Existing software architectures often struggle to achieve an optimal trade-off between these competing demands. This paper proposes SAM-ADPFL, a novel architectural framework designed to improve the engineering and management of privacy-preserving distributed machine learning systems. First, we design a geometry-aware adaptive aggregation component that dynamically reallocates aggregation weights based on local landscape properties, guiding the global model to effectively suppress model drift. Second, we propose a sharpness-guided dynamic differential privacy mechanism to handle security requirements, adaptively managing the privacy budget through refined gradient clipping and noise injection. Through empirical studies on standard benchmark datasets, we demonstrate that SAM-ADPFL exhibits superior robustness in extreme heterogeneous scenarios. By achieving a superior trade-off between system utility and privacy protection, this research contributes reusable architectural patterns and engineering practices for the design, optimization, and evaluation of privacy-preserving distributed software systems.},
DOI = {10.32604/cmc.2026.085467}
}