Homomorphic Encryption in Medical AI: The Promise and the Limits
— by Vivax
Imagine computing on patient data without ever decrypting it. That is the promise of homomorphic encryption — proven possible by Craig Gentry in 2009.
Most encryption must be undone before you can use the data. Homomorphic encryption (HE) is different: it lets you perform computations directly on encrypted data, and when you decrypt the result, it matches what you would have gotten on the plaintext. For health data, that is a tantalizing prospect — a cloud service could analyze a patient's encrypted records and return an encrypted answer, never seeing the underlying information.
The idea dates to the late 1970s, but whether arbitrary computation on encrypted data was even possible stayed open until 2009, when Craig Gentry constructed the first fully homomorphic encryption scheme in his PhD thesis — a landmark result. Practical schemes followed: BFV and BGV for exact integer arithmetic, and CKKS for approximate arithmetic on real numbers, which is the most useful for machine learning. Open libraries like Microsoft SEAL and OpenFHE now make them accessible.
The honest caveat is performance. Homomorphic operations are orders of magnitude slower and more memory-hungry than ordinary computation, and the noise that accumulates with each operation must be managed carefully. Today, HE is realistic for encrypted inference on relatively small or carefully designed models, and for specific secure-computation tasks — not for training large neural networks end to end, which remains impractical. ('Tokenization-aware' homomorphic encryption, sometimes mentioned in this context, is an emerging research direction rather than a standardized technique.)
We treat homomorphic encryption as a promising, maturing tool rather than a finished solution. For Vivax it is one layer among several — alongside differential privacy, federated learning, and confidential computing — each suited to different parts of the pipeline. Being precise about what each technique can and cannot do today is part of building medical AI that hospitals can actually trust.