CKKS ("approximate" scheme) — FHE family for real numbers. Controlled approximation noise (~10⁻¹⁰). The standard for actuarial scoring and risk models — exactly what UBI needs.

8 192 slots — Each ciphertext is a vector of 8 192 values. The 8 telematics features fill 8 slots; the rest stay zero. A single ciphertext carries the full weekly driving pattern.

Multiplicative depth 3 — How many chained multiplications are supported. Linear scoring (Mul + InnerSum + Add bias) consumes few levels.

~128 bits of security — Industry standard. Breaking the key would require ~2¹²⁸ operations.

RLWE base — Same mathematical problem behind ML-KEM/ML-DSA standardized by NIST as post-quantum cryptography.

Driver controls the data — The difference vs traditional UBI is fundamental: today, by installing the app, the driver hands over perpetual control of movement data. Under FHE, they keep the key — they can revoke the service at any moment and the insurer literally cannot read anything that was already sent.

Reduces consent fatigue — Because the data is never in cleartext on the insurer's server, the "I accept sharing X, Y, Z" screen can be drastically simplified. The user accepts "compute my discount" without having to enumerate every field used.

Raw data (stays on the phone) — In a typical week, the app captures ~10-50k GPS points + ~500k accelerometer samples (100 Hz). This raw data is highly identifying (it reveals routes, addresses, exact times) and NEVER leaves the phone.

Local feature extraction — A pipeline runs on the device itself: it computes speed from GPS deltas, detects braking events via accelerometer spikes (< -2 m/s²), classifies time of day, tags urban vs highway context via POI density. All in cleartext, locally.

Only the aggregate is encrypted — The 8 aggregated features are enough for actuarial scoring — it is the pattern Progressive Snapshot and Root Insurance have used for a decade. Raw data stays private by construction.

No persistent profile — The insurer today stores months of movement history on its own servers. Under FHE, it stores ciphertext — with no key and no ability to process it outside the authorized query context.

Breach resistant — If the insurer is hacked (and large insurers are constant targets), the attacker walks away with just ciphertext. They cannot infer where any driver has driven.

CLOUD Act compatible — Global insurers can host on US cloud without violating Brazilian LGPD. The data was never in cleartext outside the driver's phone.

Positive weights (raise risk) — hard braking (+1.5), hard acceleration (+1.2), nighttime driving (+0.8), bad weather (+0.6).

Negative weights (lower risk) — business hours (-0.2). Pattern of a driver who only drives during work hours.

Weights are public — The insurer publishes model weights so drivers can understand and audit how the discount is computed. That matters because the EU AI Act requires explainability in automated decisions affecting consumers.

Encrypted compute — The weights are public, but the feature vector is encrypted. Multiplying plaintext × ciphertext stays ciphertext. Result: encrypted score.