Compute on encrypted data without ever decrypting it. H33's FHE API ships BFV and CKKS engines with FHE-IQ auto-routing — purpose-built in Rust, not wrapped around an open-source library.
Fully Homomorphic Encryption lets you perform arbitrary computation on ciphertext. The server processes your data without ever seeing it. The result is returned encrypted — only you can decrypt it.
Think of it like a glove box: you manipulate the contents through sealed gloves. The materials are processed. The box never opens. With FHE, the "box" is mathematics, and opening it requires solving problems no computer (classical or quantum) can solve efficiently.
Traditional encryption protects data at rest and in transit. FHE protects data during computation — the last gap in the data lifecycle.
Most FHE offerings are thin API layers over Microsoft SEAL or OpenFHE. H33's FHE engines are written from scratch in Rust with production-specific optimizations:
The result: 943 μs per 32-user batch on production hardware (Graviton4 metal-48xl, 192 vCPUs).
Bit-perfect computation on encrypted integers. Ideal for biometric matching, encrypted search, and any workload requiring exact results.
Parameters: N=4096, single 56-bit modulus, t=65537. NIST Level 1 security.
Approximate arithmetic on encrypted floating-point values. Purpose-built for ML inference, statistical analysis, and neural network evaluation on encrypted data.
Supports Chebyshev polynomial activations for encrypted neural network layers.
Analyzes your workload and automatically selects the optimal FHE engine. Examines data type, precision requirements, multiplicative depth, and target latency.
Call one endpoint — FHE-IQ handles engine selection, parameter tuning, and batching strategy.
FHE computes on encrypted data. But to sign the result — to prove it was computed correctly — traditional signature systems require plaintext. This forces decryption, creating a plaintext window. The data you encrypted to protect is now exposed.
H33-74 solves this. It commits to the encrypted result directly, hashes it under SHA3-256, and signs the hash under three PQ signature families. The result: a 74-byte attestation that proves computation integrity without ever decrypting the data.
Learn more: How to Sign Encrypted Computation Without Decrypting It
| Metric | Value | Notes |
|---|---|---|
| Sustained throughput | 2,216,488 auth/sec | 30s sustained, full pipeline |
| FHE batch (32 users) | 943 μs | BFV inner product, SIMD slots |
| Per-auth latency | 42 μs | Including attest + ZKP |
| Batch attestation | 391 μs | SHA3 + Dilithium sign + verify |
| ZKP cached lookup | 0.358 μs | CacheeEngine in-process |
| Per-auth cost | ~$3.8 × 10-10 | At on-demand pricing |