A Federated Learning-Based Zero-Trust Model With Secure Dynamic Trust Evaluation and Knowledge Transfer

In next-generation networks, the increasing diversity of devices and connections, explosive data growth, and complex network threats render traditional security measures inadequate. This calls for robust, adaptive security frameworks. Zero-trust security offers robust protection through continuous verification and strict access control, while federated learning enhances data privacy and resource efficiency. Integrating these two approaches creates a resilient, adaptive, and secure network environment, meeting the intricate demands of future communication systems. In this paper, we propose FedKCSS—a Federated Learning-based zero-trust model combining Knowledge Distillation (KD) and Client Selection Strategy (CSS). FedKCSS comprises three components: Optimized Client Selection Strategy (OptCSS), Auxiliary Generator Training (AGT), and Data-free Federated Distillation (DfFD). In OptCSS, we design a dynamic trust evaluation method that continuously evaluates and adjusts client selection to enhance defense against untrusted clients. AGT involves designing a generator by using local model logits for data synthesis. DfFD is a data-free KD method that facilitates global-local model knowledge transfer, and lowering client information leakage risk without local data reliance. Experiments show that FedKCSS effectively minimizes malicious client participation in global training through dynamic trust evaluation, and improves the convergence rate by $\mathbf {8.85\%}$ and the accuracy by $\mathbf {7.09\%}$ compared with existing methods.