Scholarlink

<p>Industry 4.0 leverages information and intelligent technologies to create a highly flexible production model that supports personalized and digital products and services. However, stringent privacy protection requirements have resulted in the fragmentation and inaccessibility of production data across various companies, creating a significant data barrier. Federated learning (FL) effectively addresses privacy concerns during the data sharing process and facilitates the collaborative use of industrial production data across various institutions. However, due to the characteristics of the FL training model, attackers can implement reconstruction attacks and attribute inference attacks on the local model or global model to obtain the privacy of the data provider. Additionally, poisoning attacks by malicious users can damage the performance of FL models, thereby affecting the system’s overall utility. Many studies propose new techniques that combine privacy mechanisms and poisoning detection, but most solutions rely on outlier detection and introduce additional risks, making them impractical for real-world applications. To address these issues, we propose a privacy-preserving FL optimization method capable of resisting highly concealed poisoning attacks, FedCPA. Specifically, we protect the local model by adding eliminable noise, allowing the server to perform secure aggregation based on the fuzzy model. Furthermore, we design a ciphertext-based, two-layer architecture poisoning model detection algorithm to defend against highly concealed poisoning attacks such as Sybil attacks. Finally, the correctness of the entire calculation process is calculated and proven.</p>

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

FedCPA: A two-layer secure aggregation federated learning with lightweight privacy protection for Industry 4.0

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Jing Li,
  • Zhou Zhou,
  • Youliang Tian,
  • Chi Chen,
  • HongJun Luo,
  • LinJun He

摘要

Industry 4.0 leverages information and intelligent technologies to create a highly flexible production model that supports personalized and digital products and services. However, stringent privacy protection requirements have resulted in the fragmentation and inaccessibility of production data across various companies, creating a significant data barrier. Federated learning (FL) effectively addresses privacy concerns during the data sharing process and facilitates the collaborative use of industrial production data across various institutions. However, due to the characteristics of the FL training model, attackers can implement reconstruction attacks and attribute inference attacks on the local model or global model to obtain the privacy of the data provider. Additionally, poisoning attacks by malicious users can damage the performance of FL models, thereby affecting the system’s overall utility. Many studies propose new techniques that combine privacy mechanisms and poisoning detection, but most solutions rely on outlier detection and introduce additional risks, making them impractical for real-world applications. To address these issues, we propose a privacy-preserving FL optimization method capable of resisting highly concealed poisoning attacks, FedCPA. Specifically, we protect the local model by adding eliminable noise, allowing the server to perform secure aggregation based on the fuzzy model. Furthermore, we design a ciphertext-based, two-layer architecture poisoning model detection algorithm to defend against highly concealed poisoning attacks such as Sybil attacks. Finally, the correctness of the entire calculation process is calculated and proven.