<p>This paper proposes a Chained Encryption-based Hierarchical Secure Federated Learning (CE-HSFL) scheme designed to improve the confidentiality, integrity, and protocol compliance of model parameter aggregation in distributed federated learning (DFL), while reducing the single-point-of-failure risk in chained aggregation. The scheme leverages the Watts–Strogatz (WS) small-world network to construct the edge-layer topology, thereby enhancing node robustness. Within this architecture, CE-HSFL performs local encrypted aggregation by applying homomorphic encryption to securely aggregate the local model parameters of each edge node and its neighbors, producing encrypted local model parameters. For global updates, it employs a multi-path chained aggregation mechanism that integrates masking techniques with the small-world topology to enable distributed secure aggregation of global model parameters. To verify the correctness of parameter transmission and aggregation, zero-knowledge proof mechanisms are incorporated into both the local and global aggregation stages. CE-HSFL mitigates the impact of malicious edge nodes that submit forged, tampered, or protocol-inconsistent updates; however, data-level poisoning that remains protocol compliant is outside the direct detection scope of the cryptographic core and requires complementary anomaly detection or intrusion detection mechanisms. Theoretical analysis and experimental results indicate that CE-HSFL reduces the single-point-of-failure risk in conventional chained aggregation and improves the security and reliability of model parameter aggregation under the stated cryptographic assumptions. Furthermore, compared with the evaluated DFL baselines, CE-HSFL achieves consistent accuracy gains under the experimental settings considered in this study.</p>