Fully Homomorphic Encryption

Medical AI that never sees your data

Enclara classifies skin lesions for cancer risk using a neural network running entirely on encrypted data. Your images stay private — even from our servers.

The Problem

Medical AI demands your most sensitive data

Dermatology AI tools require patients to upload unencrypted photos of their skin to third-party servers. These images are intimate, personally identifiable medical data — and once uploaded, patients lose control over how they are stored, accessed, or shared.

Data exposure

Traditional cloud AI requires decrypted access to patient images during inference. Servers see everything.

Regulatory burden

Healthcare providers face strict GDPR and HIPAA requirements that make cloud-based AI diagnosis a compliance minefield.

Broken trust

Patients avoid digital health tools when they cannot verify their data stays private. Adoption suffers.

The Solution

Compute on encrypted data. Decrypt only results.

Enclara uses Fully Homomorphic Encryption to run a neural network directly on encrypted patient images. The server performs the entire classification without ever decrypting the data — it never sees a single pixel.

Traditional Medical AI

1Patient uploads plaintext image
2Server decrypts and processes
3Server has full access to medical data
4Trust the server, hope for the best

Enclara with FHE

1Patient encrypts image on their device
2Server runs AI on encrypted data
3Server returns encrypted results
4Only the patient can decrypt the diagnosis
Pipeline

End-to-end encrypted diagnosis

From camera to classification — your data stays encrypted at every step.

01

Capture & crop

The patient photographs a skin lesion and positions a 224×224 bounding box over the area of concern.

02

Patch, quantize & encrypt

The image is split into 49 patches (32×32 each), quantized to 5-bit precision, and encrypted on-device using the patient’s FHE secret key.

03

Encrypted inference

Each encrypted patch is sent to the server, which runs a modified VGG11 neural network entirely under FHE — no decryption occurs.

04

Decrypt results

The server returns 49 encrypted classification vectors. The client decrypts them and aggregates the final diagnosis across 7 skin cancer categories.

Technical depth

Built for real FHE inference

Not a demo — a working encrypted neural network for medical image classification.

QuantVGG11Patch

5-bit quantized VGG11 variant built with Brevitas, compiled to an FHE circuit via Concrete-ML. All MaxPool replaced with AvgPool for FHE compatibility.

Patch-based architecture

224×224 images split into a 7×7 grid of 32×32 patches. Each patch is encrypted and classified independently, then aggregated client-side.

7-class classification

Trained on HAM10000 (10,015 dermoscopic images) across akiec, bcc, bkl, df, mel, nv, and vasc — covering the major skin cancer categories.

One-time key generation

Secret key and evaluation key generated once on first launch. The evaluation key is uploaded to the server; subsequent scans reference the stored key.

Open Source

Inspect every line

Privacy claims you can't verify are just marketing. Our entire pipeline — model training, FHE compilation, client encryption, and server inference — is open source. The code is the proof.

Privacy is not a feature. It's the architecture.

Enclara proves that medical AI can be both accurate and cryptographically private. No trust required — verify it yourself.