Preserving Medical Data with Renyi Differential Privacy

The protection of medical data remains a critical concern, necessitating robust privacy-preserving techniques. Standard approaches, such as (epsilon, delta)-differential privacy (DP) are susceptible to privacy leakage during the iterative training of a deep learning model based on the composition theorem. This is a serious challenge in medical fields where deep learning models must provide guaranteed privacy preservation of medical data. On the other hand, Renyi differential privacy (RDP) has been theoretically proven to provide better privacy quantification than (epsilon, delta)-DP. Yet, there is still no research work that applies RDP to medical data. Therefore, our study investigates the potential of RDP, as a superior solution, for safeguarding medical data. We conduct a comparative analysis of the privacy and accuracy tradeoffs between RDP and non-differentially private (Non-DP) models across various optimizers. Through rigorous experimentation, our results demonstrate that RDP offers significantly improved privacy without substantial compromises in accuracy, thereby advocating for its adoption in the protection of medical data. Our findings underscore the importance of advancing privacy-preserving techniques to ensure the ethical handling of sensitive medical information in machine learning applications.