Scholarlink

Federated learning (FL) is a promising mechanism for privacy-sensitive applications, as it enables cooperation to train models among many dispersed devices without sending raw data outside. More elaborate privacy-preserving solutions are needed since the current FL principles are subject to model inversion attacks, adversarial manipulations, and data leaking. Using blockchain technology, this research examines hybrid homomorphic encryption (HHE), dynamic differential privacy (DDP), and decentralized trust verification (DTV) to study adaptive privacy-preserving federated learning (APFL). While the traditional privacy-preserving FL models are rigid in terms of privacy parameters and do not define how to integrate or change those privacy levels, APFL is built to dynamically change privacy parameters relative to immediate threats for a better balance in security and model performance. Literature review yields a clear understanding of the gap in security of current FL models compared with high-dimensional data privacy leakage and the adaptive inference attacks. Unlike most privacy-preserving approaches, in part because it adopts the adaptive privacy budget allocation, APFL endorses selective adjustments in noise levels according to relative sensitivity before applying DDP to reduce unwarranted loss of accuracy. Encrypted model updates can be made possible with a computational overhead lower than 68% that of conventional HE systems through the HHE mechanism. Moreover, a DTV powered by blockchain secures model aggregate against unauthorized changes and guarantees 99.8% data integrity. This framework was tested extensively with several privacy settings on MNIST and CIFAR-10 datasets. Evaluation results showed that APFL is superior in terms of both security and performance stability over traditional FL approaches, with accuracies of 94.2% and 98.6% on CIFAR-10 and MNIST, respectively, with a privacy budget (ε) of 1.5. Furthermore, APFL showed 24% higher computation efficiencies with a 72% loss of information leakage, making it fit for those real-world applications like smart city data analysis, financial fraud detection, and healthcare diagnostics. Fixing those significant FL deficiencies is innovative, flexible, and extraordinarily effective in ensuring the privacy of applications such as human studies.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advanced Privacy Measures for Data Sharing in Federated Learning Networks

  • 章节

  • 发表日期：

    2026年

查看全文
  • P. Manju Bala,
  • S. Usharani,
  • A. Balachandar,
  • A. Olukayode

摘要

Federated learning (FL) is a promising mechanism for privacy-sensitive applications, as it enables cooperation to train models among many dispersed devices without sending raw data outside. More elaborate privacy-preserving solutions are needed since the current FL principles are subject to model inversion attacks, adversarial manipulations, and data leaking. Using blockchain technology, this research examines hybrid homomorphic encryption (HHE), dynamic differential privacy (DDP), and decentralized trust verification (DTV) to study adaptive privacy-preserving federated learning (APFL). While the traditional privacy-preserving FL models are rigid in terms of privacy parameters and do not define how to integrate or change those privacy levels, APFL is built to dynamically change privacy parameters relative to immediate threats for a better balance in security and model performance. Literature review yields a clear understanding of the gap in security of current FL models compared with high-dimensional data privacy leakage and the adaptive inference attacks. Unlike most privacy-preserving approaches, in part because it adopts the adaptive privacy budget allocation, APFL endorses selective adjustments in noise levels according to relative sensitivity before applying DDP to reduce unwarranted loss of accuracy. Encrypted model updates can be made possible with a computational overhead lower than 68% that of conventional HE systems through the HHE mechanism. Moreover, a DTV powered by blockchain secures model aggregate against unauthorized changes and guarantees 99.8% data integrity. This framework was tested extensively with several privacy settings on MNIST and CIFAR-10 datasets. Evaluation results showed that APFL is superior in terms of both security and performance stability over traditional FL approaches, with accuracies of 94.2% and 98.6% on CIFAR-10 and MNIST, respectively, with a privacy budget (ε) of 1.5. Furthermore, APFL showed 24% higher computation efficiencies with a 72% loss of information leakage, making it fit for those real-world applications like smart city data analysis, financial fraud detection, and healthcare diagnostics. Fixing those significant FL deficiencies is innovative, flexible, and extraordinarily effective in ensuring the privacy of applications such as human studies.