Heterogeneous federated learning (HFL) is an advanced form of federated learning (FL). It relies primarily on knowledge distillation (KD) to enable knowledge transfer in heterogeneous settings. However, most existing methods rely on large volumes of data, such as high-dimensional soft labels and intermediate model features, for KD interactions across different ends, which not only incurs high communication overhead but also lacks systematic safeguards against potential privacy leakage during transmission. Furthermore, the highly non-independent and identically distributed nature of clients’ local data in HFL scenarios makes models especially vulnerable to catastrophic forgetting. To address these issues, this study proposes a dual-domain ensemble distillation-based differential HFL framework, named FedDED. Specifically, FedDED introduces a communication-efficient dual-domain ensemble distillation strategy, exchanging lightweight pseudo-labels between the server and clients to reduce communication overhead. Meanwhile, a local differential privacy strategy is introduced to perturb the pseudo-label prediction results. Furthermore, to effectively address catastrophic forgetting, a two-stage forgetting mitigation mechanism based on KL divergence regularization is proposed to enable better integration of old and new knowledge. Experimental results on four benchmark datasets demonstrate that FedDED achieves lower communication overhead and superior maximum average user accuracy (MAUA), providing a scalable and robust solution for practical distributed computing scenarios.