Fires pose serious threats to life and infrastructure, demanding fast and reliable detection. While traditional machine learning and deep learning models provide promising results, they often struggle with privacy risks, communication delays, and centralized data dependency. To address these challenges, we propose FedMA, a federated learning framework integrated with the Mayfly Optimization Algorithm (MA) and a Convolutional Neural Network (CNN) for early fire detection. MA enhances the optimization of local client models by balancing exploration and exploitation, leading to improved convergence and communication efficiency. The CNN model is trained on a balanced dataset of 28,600 fire and non-fire images sourced from Kaggle, Yandex, and Google Images. Experimental results show that FedMA achieves a detection accuracy of 97.76%, outperforming traditional Federated Averaging and Particle Swarm Optimization (FedPSO) in both accuracy and training efficiency. FedMA offers a scalable and privacy-preserving fire detection solution for real-world deployment.