Homomorphic encryption
Homomorphic encryption is a family of cryptographic techniques that allow certain computations to run on encrypted data. The result stays encrypted, and when the data owner decrypts it, it matches the result of running the same computation on the original plaintext.
What homomorphic encryption is
Homomorphic encryption, often shortened to HE, lets a program perform supported operations on encrypted values. A server might add, multiply, compare indirectly through a circuit, or evaluate a machine-learning model over ciphertext. The server sees encrypted inputs and encrypted outputs, while the data owner keeps the secret key needed for decryption. The defining promise is correctness across encryption: decrypting the computed ciphertext should give the same answer as computing directly on the original data. This makes HE a major tool in privacy-enhancing cryptography.
How encrypted evaluation works
A typical workflow starts with a client generating keys, encrypting data, and sending ciphertext to a compute service. The service runs an evaluation algorithm using public parameters and, in some schemes, evaluation keys. It returns an encrypted result. The client decrypts that result locally. The service does not receive the secret key and should not learn the plaintext from the ciphertext. The hard part is that encrypted computation changes the arithmetic, increases data size, and introduces scheme-specific limits that ordinary software does not have.
Partial, somewhat, leveled, and full
Partially homomorphic encryption supports one main operation, such as repeated additions or repeated multiplications. Somewhat homomorphic encryption supports a limited mix of operations before the ciphertext becomes too noisy. Leveled homomorphic encryption supports circuits up to a chosen depth. Fully homomorphic encryption, or FHE, supports arbitrary computable functions in principle, usually by supporting both addition and multiplication deeply enough to express general circuits. In practice, FHE still requires careful parameter selection and circuit design to make a workload feasible.
Noise, depth, and bootstrapping
Many modern HE schemes carry an internal noise term. Operations on ciphertext increase that noise, especially multiplication. If noise grows too large, decryption fails or gives the wrong answer. This is why HE systems track depth, modulus size, scaling, and other parameters. Bootstrapping is a technique that refreshes ciphertext so computation can continue. It is central to many FHE designs, but it can be expensive. Some applications avoid bootstrapping by using leveled HE for a fixed-depth computation.
Schemes and libraries
Modern HE libraries usually expose schemes tailored to different data types and workloads. BFV and BGV are often used for exact integer-style arithmetic. CKKS supports approximate arithmetic over real or complex numbers and is common in encrypted analytics and machine-learning inference. Libraries such as Microsoft SEAL, OpenFHE, HElib, PALISADE-derived projects, and commercial FHE platforms hide some low-level math, but they do not remove the need for cryptographic parameter choices, data encoding, batching, and performance testing.
Where it is useful
Homomorphic encryption is most attractive when data is sensitive, computation can be expressed as supported arithmetic, and the compute party should not see inputs or outputs. Examples include encrypted database queries, private risk scoring, health analytics, financial analytics, genomic analysis, encrypted similarity search, and privacy-preserving machine-learning inference. It is usually less suitable for arbitrary interactive software, high-throughput general computing, or workloads with lots of branching and unsupported operations. Teams often combine HE with secure multiparty computation, differential privacy, trusted execution environments, or ordinary access controls depending on the threat model.
Performance and security realities
HE can be orders of magnitude slower than plaintext computation and can produce much larger ciphertexts. Practical projects usually succeed by narrowing the computation, batching many values together, choosing the right numeric representation, and measuring latency early. Security also depends on implementation details. Parameters need to resist known lattice attacks, keys must be managed carefully, and side channels or operational mistakes can still leak information. Homomorphic encryption protects computation over encrypted data; it does not prove that a service ran the correct computation or that every surrounding system is secure.
Why it matters
Homomorphic encryption changes what cloud and cross-organization computation can mean. Instead of choosing between using data and exposing it to another party, some workflows can keep the data encrypted during processing. The idea is powerful, but its practical value comes from matching the cryptography to a narrow, important computation. HE is not a universal replacement for trust, governance, or system security. It is one specialized tool for building systems where sensitive data can be useful without being fully revealed.