Abstract
Data silos constrain cross-organizational coordination in sea–rail transport. This study integrates federated learning (FL), CKKS homomorphic encryption (HE), and blockchain (BC) into a three-layer framework for data-local collaborative learning, ciphertext aggregation, and auditable governance. Experiments use 20,000 synthetic shipment records and four clients. A same-seed ETA configuration comparison and a 500-round, eight-target evaluation show that FL+HE+BC preserves plaintext-FL fidelity: normalized RMSE is 0.0491–0.0753 and R² is 0.7578–0.8961 across six regression targets, while delay prediction reaches F1=0.9519 and AUC=0.9949. Four privacy attacks further delimit the protection boundary. For decrypted four-client aggregates, DLG success falls to 0.10, but cosine inversion remains 0.90 and gradient-property inference reaches AUC=0.7299. An external encrypted-traffic attack on high-value-cargo concentration is not significant for a single upload (best AUC=0.5595; mean permutation p=0.2468). Thus, CKKS protects update contents in transit, while decrypted aggregates and observable communication metadata remain residual risks requiring access control and traffic-shaping safeguards.
Recommended Citation
Wang, Tianyu; Zhao, Xinze; Huang, Minzhen; Lin, Xiaolei; Wang, Ying; and Zhang, Xiong, "A Privacy-Preserving Framework for Multi-Party Data Collaboration in Sea-Rail Intermodal Transport" (2026). ICEB 2026 Proceedings (Beijing, China). 2.
https://aisel.aisnet.org/iceb2026/2