Although Light Detection and Ranging (Lidar) is widely utilized in the intelligent transportation systems, the susceptibility to data degradation in snowy condition limits the application. Since rule-based methods cannot meet the requirements of real-time inference while central learning-based methods has the limitations of training data collection and privacy, this paper proposes the FedDeSnowNet, a federated de-snowing method for Lidar point clouds. In order to solve the problems arising from the feature shift as well as the client shift caused by the data heterogeneity and the communication constraints in parameter aggregation process, the FedDeSnowNet incorporates a distribution-guided aggregation mechanism as well as an ingeniously designed personalized module, both designed to enhance the training effects. FedDeSnowNet outperforms baseline methods by 14% in \(F_1\) score in experiments.