A Stagewise Framework for Implementing AI Privacy Models to Address Data Privacy and Security in Cancer Care

Abstract

Cancer patient data is not only highly sensitive but also incredibly diverse, containing genetic, clinical, and personal information. This diversity poses challenges in privacy and security, which traditional privacy models may not adequately address. This study introduces a stagewise framework for implementing AI privacy models designed to address the challenges of data privacy and security in cancer care. The framework unfolds across six stages.  The initial stage, data collection, focuses on data anonymization and masking. This step is for safeguarding personally identifiable information (PII), where sensitive details are replaced with fictional yet plausible data in preliminary datasets. As the framework progresses to the data aggregation stage, it uses federated learning and privacy-preserving record linkage (PPRL). These methods enable the integration of decentralized data from varied sources, such as different hospitals, without compromising individual identities. In the data analysis stage, differential privacy and secure multi-party computation (SMC) are employed. These techniques ensure that the analysis of aggregated data does not reveal individual patient details. Stage four of model training emphasizes using synthetic data and homomorphic encryption, necessary for training AI models with reduced privacy risks and enabling training on encrypted data. Data Sharing/Reporting, the fifth stage, includes k-anonymity and homomorphic encryption to maintain the confidentiality of shared or reported data. The final stage, Ongoing Monitoring and Updating, reiterates the continuous application of differential privacy and federated learning, essential for updating models with new data without infringing on privacy.

^{Keywords: Cancer Care, Data Privacy, Privacy Models, Security, Stagewise Framework}