Live demo · Private scoring

Encrypted credit scoring

Enter raw borrower details. The client normalizes and encrypts data, the server computes on ciphertext, and you decrypt only the final risk.

Fully Homomorphic Encryption (FHE) lets you run analytics on encrypted data without ever seeing it in the clear. Common use cases include credit scoring, healthcare analytics, and privacy-preserving ML.

Security Target 128-bit
Security Check –
Accuracy Mode Off
Input Scale 0.35

Applicant Profile

Your Result

Idle

Default Risk (reference)

–

Default Risk (private compute)

–

Credit Score (private compute)

–

Risk Tier (private compute)

–

Reference: –

Error (abs / rel)

–

Error (raw)

–

Latency

–

Error (abs / rel)

–

Bootstrap

–

Ciphertext hash

–

E2E: off
?
Session – Keygen –

Client View

  1. Normalize raw values into [-1, 1].
  2. Encrypt the normalized feature vector.
  3. Decrypt only the final risk.

Normalized Inputs

Raw (after defaults)

–

Normalized ([-1,1] * scale)

–

Server View

  1. Compute dot-product and polynomial risk on ciphertext.
  2. Return encrypted risk to the client.

Model & Timing

Scorecard (bias, scale, PDO)

–

Total latency

–

Ciphertext Snapshot

Level / Scale

–

Moduli / logN

–

c0 head

–

c1 head

–

Ciphertext bytes (head)

–

Ciphertext hash (sha256)

–

Flow Overview

  1. Client normalizes raw inputs (bounds to [-1, 1]).
  2. Client encrypts the vector using CKKS.
  3. Server runs dot-product and polynomial risk on ciphertext.
  4. Server sends encrypted risk back.
  5. Client decrypts and maps to a tier (and optional score).

Pipeline Trace

  • Normalize inputs
  • Encrypt features
  • Compute on ciphertext
  • Decrypt result

Client

Normalize → Encrypt

Private data stays local
→

Server

Compute on ciphertext

No raw access
→

Client

Decrypt → Score

Result only

Security Note

The server returns an encrypted risk signal. The client decrypts and maps it to a tier (and optionally a score) without exposing raw inputs.

Top Contributors

Largest |weight × normalized| terms

    Bootstrap Overview

    1. Coeff → Slot (FFT/linear transform)
    2. Mod Reduce (polynomial)
    3. Slot → Coeff (inverse transform)
    4. Normalize scale/level

    Bootstrap Status

    –

    Scale (runtime / target) ?

    –

    Live scale shows the real signal amplitude before refresh.

    Bootstrap latency ?

    –

    Scale ratio (runtime / target) ?

    –

    Closer to 1.0 = cleaner refresh and tighter accuracy.

    Pipeline steps

    –

    Rotation steps

    –

    Accuracy Contract

    Mod-reduce range (tanh input bound)

    –

    Normalized input bound (+/- inputScale)

    –

    Expected approx bits

    –

    –

    Presets

    Range selector (restart required)

    Preset command

    –

    Choose a range and restart the server with the env var.