Publication

Abstract : Handling sensitive medical data is critical in health-care, particularly with the rise of artificial intelligence applications. This study addresses these challenges by proposing a teacher-student framework employing differential privacy (DP) to protect data while maintaining model performance. Using Laplacian noise, predictions from teacher models are anonymized before aggregation to create student labels. The approach di- verges from traditional Federated Learning (FL) by employing the Private Aggregation of Teacher Ensembles (PATE) methodology, specifically tailored for medical imaging datasets, such as COVID-19 CT scans. Experiments demonstrate a privacy- performance trade-off, with accuracies ranging from 72% to 85%, depending on the noise level. These findings underscore the framework’s efficacy in balancing robust privacy preservation with utility. The study offers a scalable and secure method for privacy-critical healthcare applications, paving the way for reliable AI systems that adhere to stringent data protection standards. This work advances the practical integration of privacy in sensitive medical data analysis.