A Comprehensive Study on Encrypted Medical Image Inference Using AES, RSA, and Homomorphic Encryption Using AES, RSA, and Homomorphic Encryption

Artificial intelligence-based medical diagnostics is of great concern when it comes to data privacy, such as medical imaging. We propose to use symmetric (AES), asymmetric (RSA), and homomorphic encryption (HE) to create a secure AI diagnostic pipeline where the data is encrypted at all steps to preserve privacy, including at rest, in-transit and inference. We use TenSEAL when performs encrypted inference and PyCryptodome to perform cryptography tasks, and carry out experiments to measure the system performance in terms of accuracy, latency, throughput, and inference attack resistance. Our findings make it clear that a substantial level of such security may be achieved with very little performance overhead, which creates a viable long-term solution to privacy-preserving medical AI.