Federated Learning-Driven Privacy-Preserving and Security Defense for Cloud-Edge Computing: A Hierarchical Collaborative Framework

Abstract

Cloud-edge computing integrates the advantages of cloud computing‘s powerful computing capacity and edge computing‘s low-latency response, which has become the core support for data-intensive applications such as smart cities and industrial Internet of Things. However, the massive distributed data generated at the edge contains a large amount of sensitive information, and the direct transmission of data to the cloud for centralized processing faces severe privacy leakage risks. Meanwhile, the open access characteristics of edge nodes make cloud-edge systems vulnerable to various malicious attacks, which seriously threatens the security and reliability of the system. Federated Learning (FL) enables multiple participants to train models collaboratively without sharing original data, which provides an effective technical means to solve the contradiction between data sharing and privacy protection in cloud-edge computing. This study proposes a Federated Learning-Driven Hierarchical Cloud-Edge Collaborative Privacy-Preserving and Security Defense Framework (FL-HCPS). The framework adopts a two-level federated learning architecture (edge-level horizontal federation and cloud-edge vertical federation) to realize collaborative training of security models while protecting data privacy. A privacy-enhanced federated learning algorithm based on differential privacy and homomorphic encryption is designed to resist data inference attacks and model inversion attacks. In addition, an attack-aware adaptive defense mechanism is integrated to dynamically adjust defense strategies according to the type and intensity of attacks. Experimental evaluations based on two real-world datasets (EdgeIIoTset and CSE-CIC-IDS2018) show that the FL-HCPS framework achieves an average attack detection accuracy of 96.8% for common cloud-edge attacks (such as DDoS, data tampering, and model poisoning), while the data privacy leakage risk is reduced by 78.3% compared with the traditional centralized framework. The communication overhead of the framework is only 23.5% of the horizontal federated learning framework, and the model training time is shortened by 41.2%. The research results indicate that the FL-HCPS framework can effectively balance the requirements of privacy protection, security defense, and computing efficiency in cloud-edge computing, providing a new technical solution for the secure and privacy-preserving operation of cloud-edge integrated systems.