Federated learning has emerged as a promising paradigm for privacy-preserving distributed machine learning. However, Non-IID (non-independent and identically distributed) data distributions and privacy concerns remain significant challenges that hinder its broader applicability. Existing solutions typically address these challenges separately, often leading to suboptimal model performance. In this paper, we propose FedCKD-ALDP, a novel federated learning framework that integrates dual optimization techniques to simultaneously address data heterogeneity and privacy leakage. Our framework combines Clustered knowledge distillation (CKD) and adaptive local differential privacy (ALDP). CKD clusters clients based on data similarity, reducing the impact of Non-IID data distributions, while knowledge distillation within clusters enhances model performance. The ALDP mechanism dynamically adjusts perturbation levels according to layer-wise weight distributions, offering a flexible and efficient privacy-preserving solution. Experimental results on benchmark datasets—FMNIST, CIFAR10, and CIFAR100 demonstrate that FedCKD-ALDP outperforms existing methods, improving global model accuracy by up to 15.8% in Non-IID scenarios, while maintaining strong privacy guarantees with \(\varepsilon \) -local differential privacy. This framework provides a balanced solution for federated learning, optimizing both privacy protection and model performance in heterogeneous environments.