BenchmarksStack Ranking
H33-VaultH33-ShareH33-128H33-CKKSH33-256H33-BFV32H33-FHE-IQFHE OverviewZK LookupsBiometricsH33-3-KeyH33-MPCPQC ArchitecturePQ Video
APIsPricingTokenDocsWhite PaperBlogAboutSecurity Demo
Log InGet API Key

Every bank fights fraud alone. The fraudsters don’t.

Organized fraud rings operate across institutions simultaneously. But banks can’t share intelligence because it means sharing customer data. So each bank sees only its own slice — catching 20% of attacks while 80% slip through the gaps between institutions.

H33-Share changes the math.

See How It Works

H33-Share is cross-bank encrypted fraud intelligence. Banks contribute fraud signals encrypted with FHE. H33 computes joint fraud scores on the encrypted data using homomorphic operations. Neither bank ever sees the other’s data. The aggregate score is the only thing that gets decrypted.

Share intelligence. Share nothing else.

Five privacy layers between raw signals and fraud scores.

Layer 1 — FHE Encryption
Signals Encrypted Before They Leave the Bank
Each bank encrypts its fraud signals — velocity anomalies, amount spikes, device fingerprints, geo mismatches — using H33’s BFV public key. The plaintext signal values never leave the bank’s environment. H33 receives only ciphertexts that are computationally indistinguishable from random noise.
Each bank encrypts its fraud signals — velocity anomalies, amount spikes, device fingerprints, geo mismatches — using H33’s BFV public key. The plaintext signal values never leave the bank’s environment. H33 receives only ciphertexts that are computationally indistinguishable from random noise.
Layer 2 — Secure Aggregation
Pairwise Kyber Masking Before FHE
Before FHE encryption, banks apply pairwise masks using Kyber key encapsulation. Each pair of banks generates cancelling masks — so even if H33’s FHE keys were compromised, individual bank contributions remain hidden. The masks cancel in the aggregate. Defense-in-depth.
Before FHE encryption, banks apply pairwise masks using Kyber key encapsulation. Each pair of banks generates cancelling masks — so even if H33’s FHE keys were compromised, individual bank contributions remain hidden. The masks cancel in the aggregate. Defense-in-depth.
Layer 3 — Homomorphic Scoring
Joint Fraud Scores Computed on Encrypted Data
H33 multiplies encrypted signal strengths by encrypted confidences (ct×ct), applies bank weight factors (ct×pt), and accumulates across all contributing banks (ct+ct) — all without decryption. The result is an encrypted aggregate fraud score that incorporates intelligence from every bank in the consortium.
H33 multiplies encrypted signal strengths by encrypted confidences (ct×ct), applies bank weight factors (ct×pt), and accumulates across all contributing banks (ct+ct) — all without decryption. The result is an encrypted aggregate fraud score that incorporates intelligence from every bank in the consortium.
Layer 4 — Differential Privacy
Laplace Noise on Aggregate Output
After decryption, calibrated Laplace noise is added to the aggregate score. This provides mathematical differential privacy guarantees — no individual bank’s contribution can be reverse-engineered from the final output, even by an adversary with unlimited compute.
After decryption, calibrated Laplace noise is added to the aggregate score. This provides mathematical differential privacy guarantees — no individual bank’s contribution can be reverse-engineered from the final output, even by an adversary with unlimited compute.
Layer 5 — Post-Quantum Attestation
SHA3-256 Commitment + Dilithium Signature
Every fraud score carries a SHA3-256 computation commitment binding the inputs, the operation, and the output. The entire result is then signed with CRYSTALS-Dilithium ML-DSA-65 — a post-quantum signature that no future computer can forge. Banks can independently verify every score H33 produces.
Every fraud score carries a SHA3-256 computation commitment binding the inputs, the operation, and the output. The entire result is then signed with CRYSTALS-Dilithium ML-DSA-65 — a post-quantum signature that no future computer can forge. Banks can independently verify every score H33 produces.

Real-time fraud scoring across encrypted inputs from multiple banks.

H33-Share processes 32 entities per FHE ciphertext using SIMD batching on BFV lattice encryption. A full scoring round — 8 fraud categories, 3 contributing banks, homomorphic multiply + accumulate + decrypt + attest — completes in under 1.3 milliseconds per batch.

For a consortium of 10 banks tracking 1 million entities, that’s a complete cross-bank fraud intelligence refresh in under 41 seconds. Every entity scored against signals from every bank, without any bank seeing another’s data.

<1.3ms
per 32-entity batch — 8 categories, 3+ banks, fully encrypted

Faster than a database round-trip. Five cryptographic privacy layers deep.

Cross-Bank Intelligence Flow

A
Bank A
B
Bank B
C
Bank C
H33
Encrypted Compute
Entity #1
—
Entity #2
—
Entity #3
—
ZERO DATA VISIBILITY — ONLY AGGREGATES DECRYPTED

Eight encrypted signal categories. One unified score.

Velocity Anomalies
Rapid transaction bursts
Catches card-testing attacks that span multiple issuers simultaneously.
FHE-encrypted counters
Amount Anomalies
Unusual transaction sizing
Detects structuring patterns invisible to any single institution.
Homomorphic comparison
Geo Anomalies
Impossible travel patterns
Cross-bank geo correlation catches account takeover across institutions.
Encrypted coordinates
Device Anomalies
Fingerprint clustering
Same device hitting multiple banks — visible only in aggregate.
Hashed device IDs
Account Takeover
Credential compromise signals
Stolen credentials tested across institutions detected collectively.
Cross-bank correlation
Synthetic Identity
Fabricated identity detection
Synthetic IDs that pass any single bank’s KYC fail the consortium check.
Multi-bank KYC fusion
Known Patterns
Historical fraud signatures
Federated model training across banks — Byzantine-resilient aggregation.
TrimmedMean + DP
Money Mule
Layering and placement
Fund movement across accounts at different banks traced in encrypted form.
Encrypted graph analysis

Contributing signals is free. Score queries cost units.

Same H33 unit system as Auth and Vault. Volume drives discounts at every level. Signal contributions are always free.

Share-0
5 units per query
2 signal categories (Velocity + Amount). Dilithium-signed results. SHA3-256 commitment. Basic differential privacy.
<25K units$0.30/q
250K-2.5M$0.125/q
25M+$0.03/q
Share-1
10 units per query
All 8 categories. Per-category breakdown. FHE velocity tracking. Calibrated differential privacy (ε=2.0).
<25K units$0.60/q
250K-2.5M$0.25/q
25M+$0.06/q
Share-2
20 units per query
Share-1 + pairwise Kyber masking (defense-in-depth). Federated model training. Byzantine-resilient aggregation.
<25K units$1.20/q
250K-2.5M$0.50/q
25M+$0.12/q
Share-3
35 units per query
Share-2 + custom model weights. Dedicated FHE compute. 99.99% SLA. Private consortium deployment. Priority scoring.
<25K units$2.10/q
250K-2.5M$0.875/q
25M+$0.21/q

Volume Unit Pricing

Monthly Volume$/UnitShare-0 (5u)Share-1 (10u)Share-2 (20u)Share-3 (35u)
Under 25K units$0.060$0.30$0.60$1.20$2.10
25K – 250K$0.040$0.20$0.40$0.80$1.40
250K – 2.5M$0.025$0.125$0.25$0.50$0.875
2.5M – 25M$0.012$0.06$0.12$0.24$0.42
25M+ units$0.006$0.03$0.06$0.12$0.21

Share-2 at 25M+ = $0.12/query — full FHE + Kyber + Dilithium + differential privacy

Per-Operation Costs

OperationUnitsNotes
Signal encrypt + ingest0 (free)Incentivizes sharing
Homomorphic accumulate1Per category per entity
Score decrypt + DP noise2Final aggregate only
Dilithium attestation0Included with every query
SHA3-256 commitment0Included with every query
Kyber secure agg round5Share-2+ only
Federated model update3Share-2+ only
10K queries/mo
Units Used
100K
Monthly Spend
$2,500
Per Query
$0.25
H33-ShareConsortium DBData Broker
Share-2 per query (volume)$0.12$0.50–2.00$1.00–5.00
Data visibility to operatorZero — FHE encryptedFull plaintextFull plaintext
Post-quantum signaturesDilithium (FIPS 204)——
Homomorphic computationBFV lattice (N=4096)——
Differential privacyLaplace (ε=2.0)——
Regulatory riskNone — no data sharingHigh — data poolingHigh — CCPA/GDPR

No data pooling. No plaintext sharing. No regulatory exposure.

Pure H33 stack: BFV FHE + Kyber secure aggregation + Dilithium signatures + differential privacy.

Join the Consortium Read the API Docs