Federated learning is prone to privacy leakage issues. Current research on secure aggregation protocols can protect user privacy during the federated learning process. However, as the number of participants and the volume of data increase, the communication overhead and computation time grow polynomially. Therefore, traditional secure aggregation protocols are challenging to apply in resource-constrained environments. Recent works have customized secure aggregation protocols based on Rand-K and Top-K sparsification, which have significantly reduced communication and computation overhead. Nevertheless, there is still substantial room for improvement. This paper builds on Top-K sparsification by integrating quantization and customizing a secure aggregation protocol SecSQ which is based on homomorphic encryption. This approach further reduces the amount of encryption and decryption operations, thereby achieving lower communication and computation costs. Additionally, this paper designs an oblivious random grouping protocol that randomizes group assignments and designates temporary clients for aggregation in each iteration. This creates a dynamically changing pyramid structure within the federated learning framework, which not only protects user privacy but also distributes the server’s computational and communication load. Experiments demonstrate that the proposed method reduces global communication volume by 70% and total computation time by 50% compared to the latest Top-K sparsification research.