Scholarlink

Federated Learning (FL) enables decentralized model training while preserving data locality, but model parameter transmission is vulnerable to privacy attacks. Existing techniques like Differential Privacy (DP), Homomorphic Encryption (HE), and Secure Multi-Party Computation (SMC) face trade-offs: DP degrades model utility due to noise, while HE/SMC incurs high computational costs. We propose KD-IBMRKE-PPFL, a dual-phase framework integrating Knowledge Distillation (KD), Identity-Based Multi-Receiver Key Encapsulation (IBMRKE), and DP to address these issues. During aggregation, a Bernoulli-based DP mechanism perturbs client-predicted labels with controlled noise to protect sensitive data. In dissemination, IBMRKE ensures secure transmission of the global model through key encapsulation, restricting decryption to authorized clients. KD helps mitigate challenges from model heterogeneity by transferring knowledge from a teacher model to client models with different architectures, enabling efficient aggregation. Experimental results on MNIST, Fashion-MNIST, and Chest-Xray datasets demonstrate that KD-IBMRKE-PPFL achieves 94%, 81%, and 75.2% accuracy under stringent privacy budgets ( \( \epsilon =2 \) ) with only 1–3% performance loss, outperforming FedAvg, FedSGD, and FedAdam by 8–12%. The framework balances security, efficiency, and utility, offering a practical solution for privacy-preserving FL.

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

KD-IBMRKE-PPFL: A Privacy-Preserving Federated Learning Framework Integrating Knowledge Distillation and Identity-Based Multi-receiver Key Encapsulation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yuan Li,
  • Changji Wang,
  • Shiwen Hu

摘要

Federated Learning (FL) enables decentralized model training while preserving data locality, but model parameter transmission is vulnerable to privacy attacks. Existing techniques like Differential Privacy (DP), Homomorphic Encryption (HE), and Secure Multi-Party Computation (SMC) face trade-offs: DP degrades model utility due to noise, while HE/SMC incurs high computational costs. We propose KD-IBMRKE-PPFL, a dual-phase framework integrating Knowledge Distillation (KD), Identity-Based Multi-Receiver Key Encapsulation (IBMRKE), and DP to address these issues. During aggregation, a Bernoulli-based DP mechanism perturbs client-predicted labels with controlled noise to protect sensitive data. In dissemination, IBMRKE ensures secure transmission of the global model through key encapsulation, restricting decryption to authorized clients. KD helps mitigate challenges from model heterogeneity by transferring knowledge from a teacher model to client models with different architectures, enabling efficient aggregation. Experimental results on MNIST, Fashion-MNIST, and Chest-Xray datasets demonstrate that KD-IBMRKE-PPFL achieves 94%, 81%, and 75.2% accuracy under stringent privacy budgets ( \( \epsilon =2 \) ) with only 1–3% performance loss, outperforming FedAvg, FedSGD, and FedAdam by 8–12%. The framework balances security, efficiency, and utility, offering a practical solution for privacy-preserving FL.