With the rapid advancements in technology and data analytics, transportation agencies are keen on adopting innovative solutions to enhance their infrastructure management plans. Given this, condition prediction models have received more attention due to their impact on the effective management of inspection resources. Machine learning (ML) algorithms are typically employed in the development of these prediction models. Traditional centralized ML models often face challenges related to data privacy, transferability, and integration from multiple sources. This paper proposed an innovative approach for infrastructure condition prediction models by leveraging Federated Learning (FL), a decentralized ML paradigm. To illustrate the proposed approach, we selected culverts in Utah as a case study. In addition to the Utah culvert inventory, we obtained further data from inventories in five other states of the US. We presented a comparative analysis of two federated models—Federated Proximal and Federated Averaging—as well as a centralized ML model. Our findings highlight the efficacy of the proposed FL-based models in enhancing prediction accuracy while ensuring data privacy and reducing data transmission overheads. Furthermore, Utah was able to leverage insights from other states’ inventories through Federated Learning since it faced a data deficiency regarding culvert inspection records.