Agriculture is paramount in the world feeding billions and is vital to the food supply and the economy and stability of the world, yet little or no attention has been paid to the salt that poisons our global agricultural land. In the case of major crops, maize is particularly important, but its productivity is commonly compromised by foliar diseases, which in turn can cause dramatic loss of yield. There is a critical need for early detection of such diseases in field or farm situations in order to enable timely response, however scale in such response and identification is non-trivial, especially when decentralized per farm or region reporting is enforced by privacy regulations. To tackle this, we introduce an FL-based privacy-preserving deep learning strategy which allows training the models in a decentralized manner across the distributed data sources without consolidating the sensitive training data on a centralized server. The model is constructed using the architecture of the EfficientNetV2-S convolutional neural network (CNN), which was augmented with squeeze and excitation (SE) blocks to perform better feature representation for fine-grained disease recognition. Through this framework, multiple institutions or farms can be involved in training a disease detection model collaboratively without sharing their data, ensuring data privacy. Generally, the method integrates an improved CNN architecture with FL to address the simultaneous problem of accurate disease detection on maize leaf and scalable privacy preservation.