Complete reference for every cryptographic parameter used across H33's FHE engines, post-quantum signature families, key encapsulation, hash functions, and TFHE gate operations. All parameters are chosen to meet or exceed NIST security level targets.
The Brakerski/Fan-Vercauteren (BFV) scheme is H33's primary integer-arithmetic FHE engine. Parameters are selected for each security tier to balance security level, noise budget, and throughput.
BFV operates on polynomial rings R_Q = Z_Q[X]/(X^N + 1), where N is the polynomial degree, Q is the ciphertext modulus, and t is the plaintext modulus. Security depends primarily on the ratio N/log2(Q) -- larger N and smaller Q yield higher security. H33 provides five tiers, each targeting a specific use case and NIST security level.
The production pipeline uses H33-128 (biometric_fast mode) for high-throughput authentication, achieving 2,209,429 auth/sec sustained on Graviton4 hardware. Higher tiers are available for applications requiring NIST Level 5 security or larger computation depth.
| Tier | N (Degree) | Q (Ciphertext Mod) | t (Plaintext Mod) | NIST Level | Use Case |
|---|---|---|---|---|---|
| H0 (Demo) | 1024 | 27-bit | 257 | Below Level 1 | Testing, development, demos |
| H1 (Light) | 2048 | 40-bit | 65537 | Level 1 | Low-sensitivity workloads |
| H33 (Production) | 4096 | 56-bit | 65537 | Level 1-3 | Production auth pipeline (biometric_fast) |
| H2 (Enhanced) | 8192 | 109-bit (2 limbs) | 65537 | Level 3 | Deeper computation, regulated workloads |
| H-256 | 16384 | 218-bit (4 limbs) | 65537 | Level 5 | Maximum security, deep multiplication chains |
Production pipeline uses H33 tier (N=4096, single 56-bit modulus). Montgomery radix-4 arithmetic with Harvey lazy reduction. biometric_fast() mode: single Q limb, no RNS decomposition overhead. FHE batch latency: 943 microseconds for 32 users.
The Cheon-Kim-Kim-Song (CKKS) scheme enables approximate arithmetic on encrypted floating-point data. H33-CKKS is used for machine learning inference, statistical analysis, and any workload requiring real-number computation on encrypted data.
CKKS encodes complex numbers into polynomial coefficients with a scaling factor (Delta). Each multiplication consumes one level of the modulus chain, so the number of available levels determines the maximum multiplicative depth. Rescaling after each multiplication maintains precision by reducing the ciphertext modulus.
H33-CKKS supports configurable precision from 20-bit to 50-bit mantissa accuracy, depending on the tier and depth requirements. The H33-CKKS engine page provides detailed benchmarks for each tier on Graviton4 hardware.
| Tier | N | Max Depth | Scale (bits) | Precision | NIST Level |
|---|---|---|---|---|---|
| CKKS-Light | 4096 | 3 | 30 | ~20 bits | Level 1 |
| CKKS-Standard | 8192 | 7 | 40 | ~30 bits | Level 3 |
| CKKS-Deep | 16384 | 15 | 50 | ~40 bits | Level 5 |
| CKKS-Max | 32768 | 30+ | 50 | ~50 bits | Level 5 |
Graviton4 CKKS metal benchmarks: 61ms multiply, 333ms dot product, 14 ops/sec batch throughput. CRT-consistent RLWE key generation: key-switch (a,e) are integers reduced per modulus, not independently sampled.
Torus Fully Homomorphic Encryption (TFHE) enables gate-by-gate Boolean computation on encrypted bits. H33-TFHE supports bootstrapping after every gate, maintaining a constant noise level for arbitrary-depth circuits.
TFHE operates on the torus T = R/Z. Each gate operation (AND, OR, XOR, NAND, NOR, XNOR, NOT) produces a ciphertext with bounded noise. Programmable Bootstrapping (PBS) resets the noise to a fixed level while applying a lookup table, enabling unbounded computation depth. H33-TFHE includes multi-bit optimizations for 8-bit, 16-bit, 32-bit, and 64-bit integer operations.
| Gate Type | Noise Budget | 96-Channel TPS | PBS | Notes |
|---|---|---|---|---|
| AND / NAND | Consumes ~1 bit | 768 (8-bit GT) | After each gate | Universal gate set |
| OR / NOR | Consumes ~1 bit | 768 (8-bit GT) | After each gate | Universal gate set |
| XOR / XNOR | Consumes ~1 bit | 768 (8-bit GT) | After each gate | Parity / comparison |
| 8-bit Greater-Than | Full PBS chain | 768 | Per-gate | Comparison circuit |
| 16-bit Greater-Than | Full PBS chain | 372 | Per-gate | Extended precision |
| 32-bit Greater-Than | Full PBS chain | 182 | Per-gate | Integer comparison |
| 64-bit Greater-Than | Full PBS chain | 91 | Per-gate | Large integer comparison |
| 16-bit Equality | Full PBS chain | 769 | Per-gate | Encrypted matching |
GPU Multi-bit TFHE: 1,129 TPS on A10G (1.63x over v4 691 TPS). 1.0% noise budget matches CPU reference implementation. Three correctness bugs fixed via iteration-count diagnostic sweep.
Module-Lattice-based Digital Signature Algorithm. H33's primary signature family, based on the hardness of the Module Learning With Errors (MLWE) problem.
ML-DSA-65 (formerly CRYSTALS-Dilithium Level 3) is the first of H33's three post-quantum signature families. It provides NIST Security Level 3, meaning it offers at least as much security against quantum attacks as AES-192 does against classical attacks. ML-DSA is the mandatory signature algorithm in NIST FIPS 204 and is expected to see the broadest adoption across government and industry.
| Parameter | Value |
|---|---|
| NIST Standard | FIPS 204 |
| Security Level | NIST Level 3 (128-bit quantum security) |
| Hardness Assumption | Module-LWE (MLWE) |
| Public Key Size | 1,952 bytes |
| Signature Size | 3,293 bytes |
| Secret Key Size | 4,000 bytes |
| Signing Speed (Graviton4) | ~35 microseconds per signature |
| Verification Speed | ~12 microseconds |
| Module Rank (k, l) | (6, 5) |
| Challenge Weight (tau) | 49 |
Module-Lattice-based Key Encapsulation Mechanism. Used for all key exchange operations in H33's post-quantum infrastructure.
ML-KEM-1024 (formerly CRYSTALS-Kyber-1024) provides NIST Security Level 5 key encapsulation, ensuring that key exchange remains secure even against the most powerful quantum adversaries. H33 uses ML-KEM-1024 for all session key establishment, API key exchange, and inter-service key agreement within the post-quantum API infrastructure.
| Parameter | Value |
|---|---|
| NIST Standard | FIPS 203 |
| Security Level | NIST Level 5 (256-bit quantum security) |
| Hardness Assumption | Module-LWE (MLWE) |
| Public Key Size | 1,568 bytes |
| Ciphertext Size | 1,568 bytes |
| Shared Secret Size | 32 bytes |
| Module Rank (k) | 4 |
| Encapsulation Speed | Sub-millisecond on Graviton4 |
| Decapsulation Speed | Sub-millisecond on Graviton4 |
Fast Fourier Lattice-based Compact Signatures over NTRU. H33's second signature family, providing an independent hardness assumption from ML-DSA.
FALCON-512 is based on the NTRU lattice problem, which is mathematically independent from the Module-LWE problem underlying ML-DSA. This independence is critical for H33's defense-in-depth strategy: even if a future cryptanalytic breakthrough compromises MLWE, FALCON's NTRU-based security remains intact. FALCON produces the most compact signatures of H33's three families, making it particularly valuable for bandwidth-constrained applications.
| Parameter | Value |
|---|---|
| NIST Standard | Selected (Round 3), FIPS pending |
| Security Level | NIST Level 1 |
| Hardness Assumption | Short Integer Solution over NTRU lattices |
| Public Key Size | 897 bytes |
| Signature Size | ~666 bytes (variable, Gaussian sampling) |
| Secret Key Size | 1,281 bytes |
| Degree (N) | 512 |
| Modulus (q) | 12289 |
| Signing | Requires floating-point (discrete Gaussian sampling) |
FALCON-512 is NIST Level 1. The upgrade to FALCON-1024 (Level 5) is pending Graviton4 A/B testing. Until that upgrade ships, the three-key bundle is bounded at Level 1 security. The correct marketing claim is "three independent mathematical hardness assumptions," not "three Level 3 families."
Stateless Hash-based Digital Signature Algorithm (SPHINCS+). H33's third signature family, based solely on the security of hash functions -- the most conservative possible post-quantum assumption.
SLH-DSA (formerly SPHINCS+) is unique among H33's signature families because its security depends only on the properties of the underlying hash function (SHA-256 in this configuration), not on any structured mathematical problem. If hash functions remain secure, SLH-DSA remains secure -- regardless of advances in lattice cryptanalysis or quantum computing. This makes it the ultimate fallback in H33's defense-in-depth strategy.
The "-128f" variant is optimized for fast signing at NIST Level 1 security. The "f" designation indicates the "fast" parameter set, which trades larger signature sizes for faster signing speed. H33 uses the SHA2 variant for compatibility with the OpenSSL SHA-256 acceleration shim deployed on Graviton4 (4.33x speedup vs. generic implementation).
| Parameter | Value |
|---|---|
| NIST Standard | FIPS 205 |
| Security Level | NIST Level 1 |
| Hardness Assumption | Collision resistance + second-preimage resistance of SHA-256 |
| Public Key Size | 32 bytes |
| Signature Size | 17,088 bytes |
| Secret Key Size | 64 bytes |
| Hash Function | SHA-256 |
| Tree Height (h) | 66 (hypertree) |
| WOTS+ Parameter (w) | 16 |
H33 uses SHA3-256 (Keccak) with domain-separated contexts for all internal hashing operations. Domain separation ensures that identical inputs used in different contexts produce different hash outputs, preventing cross-context collision attacks.
Every hash computation in H33's infrastructure includes a domain separator prefix that identifies the context of the operation. This prevents an attacker from using a hash computed in one context (e.g., an attestation transcript) as a valid hash in another context (e.g., a key derivation). Domain separators are frozen as part of H33's protocol stability guarantee and cannot be modified without a major version increment.
| Domain | Separator Prefix | Usage |
|---|---|---|
| Attestation Transcript | H33-ATTEST-V1 | H33-74 attestation transcript binding |
| Key Derivation | H33-KDF-V1 | Session key and sub-key derivation |
| Chain State | H33-CHAIN-V1 | Auth event chain state transitions |
| Replay Integrity | H33-REPLAY-V1 | Governance replay classification |
| Conformance Vector | H33-CONFORM-V1 | HATS conformance vector derivation |
| Tombstone | H33-TOMBSTONE-V1 | Cryptographic deletion proofs |
| Anchor | H33-ANCHOR-V1 | External chain (Bitcoin/Solana) anchoring |
Measured on Graviton4 c8g.metal-48xl (192 vCPUs, 371 GiB) with system allocator. All numbers are sustained, not burst.
| Algorithm | Operation | Latency | Throughput |
|---|---|---|---|
| BFV (H33 tier) | 32-user batch encrypt + inner product | 943 microseconds | ~33,900 batches/sec |
| ML-DSA-65 | Sign + Verify (batch attestation) | 391 microseconds (batch of 32) | ~2.2M auth/sec sustained |
| FALCON-512 | Sign | Sub-millisecond | Part of three-key bundle |
| SLH-DSA-SHA2-128f | Sign | Dominated by hypertree traversal | Part of three-key bundle |
| ML-KEM-1024 | Encapsulate + Decapsulate | Sub-millisecond | Session establishment |
| SHA3-256 | Hash (domain-separated) | ~96 nanoseconds (w/ OpenSSL shim) | ~10M hashes/sec |
| ZK-STARK (cached) | Lookup verification | 0.358 microseconds | ~2.8M lookups/sec |
| CKKS (Standard) | Encrypted multiply | 61 milliseconds | 14 ops/sec batch |
| TFHE (8-bit GT) | Greater-than comparison | ~1.3 milliseconds | 768 TPS (96-channel) |
How H33's algorithms map to NIST's defined security levels. Each level represents the computational effort required to break the algorithm, benchmarked against the equivalent classical symmetric cipher.
| NIST Level | Classical Equivalent | H33 Algorithms at This Level |
|---|---|---|
| Level 1 | AES-128 | FALCON-512, SLH-DSA-SHA2-128f, BFV H33-tier, CKKS-Light |
| Level 2 | SHA-256 collision | -- |
| Level 3 | AES-192 | ML-DSA-65, BFV H2-tier, CKKS-Standard |
| Level 4 | SHA-384 collision | -- |
| Level 5 | AES-256 | ML-KEM-1024, BFV H-256 tier, CKKS-Deep, CKKS-Max |
The three-key signature bundle (ML-DSA-65 + FALCON-512 + SLH-DSA-SHA2-128f) is bounded at NIST Level 1 until the FALCON-1024 and SLH-DSA-SHA2-192f upgrades complete Graviton4 A/B testing. The correct claim is "three independent mathematical hardness assumptions" -- MLWE, NTRU, and stateless hash functions.
These parameters drive real production workloads. Test them yourself via the live API, or explore the engine pages for integration guides.