This work presents the use of Federated Learning (FL) in cloud environments to enhance performance and privacy in machine learning applications. The main challenge addressed is delivering a secure, decentralized model training with user privacy preservation. With data protection becoming increasingly important, FL presents a way to train models on distributed devices without exposing sensitive data to central servers. This study utilizes the Federated EMNIST dataset to simulate FL environments, employing algorithms that enable privacy-preserving model updates. The results show that FL can achieve 6% higher accuracy than traditional centralized learning, with a final accuracy of 89% at the fifth epoch compared to 83% in traditional approaches. FL reduces the training time by 20% compared to conventional learning, thus lowering the training time from 50 s for the initial epochs to 40 s for the fifth epochs. In contrast, traditional learning reduces from 60 s to 50 s. Overall, this paper addresses the feasibility of Federated Learning as a privacy-guaranteeing, scalable alternative to cloud-based machine learning and its feasibility to enter areas handling sensitive data like finance and healthcare.