In the age of cloud computing, effectively managing and balancing large data loads across distributed servers presents an important task. In traditional centralized load balancing strategies, there are some common issues with data privacy, communication overhead, and scalability. To improve cloud load balancing, this paper presents Federated Autoencoder-Based Load Regulation (FA-LR) method which facilitates training the model locally. Our Proposed Method combines federated learning with autoencoder architectures that provide decentralized training of models on different cloud nodes. It makes sure that raw data stays local, thereby preserving privacy. The main objective of autoencoders are to compress and reconstruct load data, allowing efficient communication and precise load predictions. When comparing the FA-LR method to traditional load balancing strategies, the results show that it performs better in terms of load distribution efficiency up to 92.7%, lowers communication costs by 31%, and improves data privacy.