Background: Plant diseases cause an estimated 20–40 percent of annual crop yield losses worldwide, and most deep learning-based diagnostic systems depend on centralized data pipelines that require farmers to transmit raw crop images to remote servers, raising privacy, connectivity, and data-sovereignty concerns. 
Objectives: This study aimed to develop and evaluate LeafFed, a federated learning framework for privacy-preserving, multi-node plant disease classification that avoids transferring raw images between clients. 
Methods: LeafFed employs a RegularizedEfficientViT model (a frozen EfficientNet-B0 backbone with a regularized six-layer classifier head) trained on the 38-class PlantVillage color dataset, partitioned across three independent client nodes and optimized using the FedAvg algorithm over five communication rounds. 
Key Findings: The resulting global model achieved 92.83 percent test accuracy, 92.50 percent precision, 92.60 percent recall, and a 92.55 percent F1-score, a 5.49 percentage-point improvement over a centralized baseline (88.00 percent accuracy). Test loss decreased monotonically from 2.79 to 0.94 across rounds, and per-client analysis revealed substantial cross-node knowledge transfer, with the weakest client rising from 37.24 percent to 85.31 percent accuracy. These results show that federated learning can match or exceed centralized performance while preserving farm data privacy and sovereignty.