Smart farming technologies, particularly image-based crop disease detection systems, are significantly transforming agriculture by enabling farmers to respond promptly to emerging threats. However, deploying centralized AI models in remote or rural farming communities remains challenging due to limited internet connectivity, heterogeneous data sources, and privacy concerns. This paper presents a federated learning (FL) framework designed to support secure and privacy-preserving crop disease detection in decentralized agricultural environments. In the proposed approach, lightweight convolutional neural networks (CNNs) are trained locally on crop images captured by drones or field sensors, ensuring that sensitive data remains at the source. Instead of transmitting raw data, only model updates are shared with a central server, which aggregates them using the Federated Averaging (FedAvg) algorithm. To enhance security and model integrity, a cosine similarity-based filter is incorporated to detect and discard malicious or corrupted updates. Additionally, an attention-based heatmap module is introduced to visually highlight disease-affected regions in the crop images, improving system transparency and building user trust. Experimental evaluations on the PlantVillage dataset and real-world drone imagery of wheat and tomato crops demonstrate that the framework achieves an average 4.8% improvement in F1-score under non-IID data conditions compared to centralized models, alongside a 9.3% reduction in communication overhead at 50 nodes. This solution is well-suited for resource-constrained rural settings and contributes to scalable, intelligent, and sustainable agricultural digitization.