Homomorphic encryption

In one line

Homomorphic encryption (HE) lets a server run computations directly on ciphertext: it returns an encrypted result that only the data owner can decrypt, having learned nothing along the way.

The problem it solves

Normally, to compute on data you must decrypt it — which means the machine doing the work (a cloud server, a research partner) sees the plaintext. For genomic or claims data that's a showstopper. Homomorphic encryption breaks the rule: it keeps data encrypted even during computation, so you can outsource analysis to a party you don't trust to see the data.

How it works

• Schemes support addition and multiplication on encrypted values; chaining those builds up arbitrary computations.
• Fully homomorphic encryption (FHE) — the BGV/BFV/CKKS lineages — supports arbitrary circuits. CKKS handles the approximate real-number math that statistics and ML need.
• The owner encrypts, the server computes on ciphertext, the owner decrypts the result — and the server learned nothing.

The catch: cost

HE operations are orders of magnitude slower than plaintext, and ciphertexts are large. So practical systems use it surgically — encrypting just one sensitive scoring step — or in hybrid designs with secure enclaves and secure multi-party computation (MPC). It is not a drop-in replacement for normal computation.

Where it sits in the privacy spectrum

Think of a spectrum: TLS (encrypt in transit) → encryption at rest → federated learning (don't move the data) → differential privacy (bound what outputs reveal) → homomorphic encryption (compute without ever seeing). HE is the far, strongest — and costliest — end.

Where it shows up in digital health

• Outsourcing analysis of genomic or claims data to an untrusted cloud.
• Cross-institution statistics where even the aggregator must not see inputs.
• Encrypted inference — a model scores an encrypted record.

Today it's a research-and-niche tool gaining ground, worth knowing as the endpoint of the privacy-tech spectrum.

Common pitfalls

• Reaching for FHE first — its cost usually means a lighter tool (DP, federated, MPC) fits better; use HE where nothing else will do.
• Underestimating performance — benchmark the specific computation before committing.
• Assuming it covers everything — HE protects data in use; you still need access control and key management.

Key takeaways

• HE computes directly on ciphertext — the processor never sees plaintext.
• FHE (BGV/BFV/CKKS) supports arbitrary computation; CKKS suits ML/stats on real numbers.
• It's powerful but expensive — use surgically or in hybrid designs.
• The strongest, costliest end of the privacy spectrum that begins at TLS.