Abstarct: Abstract The rapid mutation of SARS-CoV-2 necessitates accurate and efficient variant classification to support timely public health responses. In this study, we present an advanced federated learning (FL) frxamework that enables privacypreserving and decentralized classification of SARS-CoV-2 variants without requiring direct data sharing between institutions. By leveraging distributed data analysis, our approach ensures that local datasets remain confidential while still contributing to the global model’s learning process. To further improve classification performance, we incorporate Graph Neural Networks (GNNs) into the FL architecture. GNNs are particularly effective for modeling complex relationships in protein sequence data, allowing for more accurate variant identification. This hybrid approach not only enhances the predictive power of the model but also maintains data security and integrity across multiple research centers. Our experimental results demonstrate that the proposed method achieves 93.26% accuracy in classifying SARS-CoV-2 variants, outperforming several traditional machine learning approaches. Additionally, the frxamework is designed to be scalable, making it suitable for large-scale genomic studies and real-world pandemic monitoring applications. By integrating federated learning and deep graph-baxsed representations, our model offers a robust, adaptable, and privacy-focused solution for genomic data analysis, contributing to more effective tracking of viral mutations and aiding global health initiatives.