Drug interaction prediction has been an important issue in healthcare in recent years, as it helps to ensure patient safety and treatment effectiveness. The problem, however, is that real-world medication interaction data is scarce and subject to privacy rules, making it difficult to obtain. Building reliable models using conventional centralized data sources has become challenging due to these constraints. Federated learning (FL), a privacy-preserving machine learning paradigm, has gained popularity as a solution to problems. Without requiring the sharing of sensitive, raw data, FL enables the decentralized training of models across several data sources. We provide an improved federated learning approach in this study that uses artificial datasets created for different clients to predict medication interactions. In the TensorFlow Federated (TFF) context, our method improves model generalization by utilizing deep learning methods like Batch Normalization and Dropout. We also examine how various optimization techniques, in conjunction with increased Federated Averaging, impact model performance across several training cycles. Even under strict privacy limitations, the suggested methodology provides a promising path toward safe and effective medication interaction predictions.