a technical overview

fully homomorphic encryption

Fully Homomorphic Encryption (FHE) is an advanced cryptographic technique that enables computations on encrypted data without ever decrypting it.

This means that even when data is processed by AI models, cloud services, or third parties, its confidentiality remains intact.

How FHE differs from traditional encryption?

Traditional encryption protects data at rest or in transit but require decryption for processing.

AI Inference with FHE

In Lattica we build an AI inference platform that runs AI models on encrypted queries using FHE:

Step 1
encrypt on device

A query is encrypted directly in your browser using FHE.

This ensures that Lattica never sees or stores your raw data, only its encrypted form.

Step 2
COMPUTE ON CLOUD

The AI model runs directly on the encrypted query without first decrypting it.

Lattica cannot see your input, output, or intermediate results during computation.

Step 3
receive encrypted result

The model returns an encrypted prediction.

Step 4
DECRYPT ON DEVICE

Decrypt and view the final prediction on your browser.

If FHE is so powerful, why isn't it everywhere?

While FHE is a breakthrough technology in privacy-preserving computation, several challenges have slowed its adoption:

hight computational costs

FHE operations require significantly more processing power conpared to traditional computing, making them slower and more resource-intensive.

LIMITED PRACTICAL OPTIMIZATION

Most existing FHE tools originate from applied academic research and are designed for general-purpose use, making them inefficient for real-world applications.

DIFFICULT DEPLOYMENT

Enterprises face challenges integrating FHE due to a lack of readily available cryptographic expertise.

FRAGMENTED ECOSYSTEM

The absence of standardization across the FHE stack creates compatibility issues, making integration complex and time-consuming.

What we do at Lattica

Our solution, expertise and attention are at making FHE work good at scale for neural networks inference. This means we optimize the whole stack:

PLATFORM

A system for deploying and managing neural networks, supporting both a web interface and API-based interactions. It integrates with the standard ML stack and enables dynamic compute allocation and access management. We call this FHE-as-a-Service.

QUERY CLIENT

Handles key generation, encryption, and decryption. Running across multiple platforms, from browsers to Python and mobile.

FHE INFERENCE ENGINE

The core of our backend, responsible for executing neural networks on encrypted queries. This is where we apply algorithmic optimizations tailored to both network architecture and encryption security parameters.

HEAL (Homomorphic Encryption Abstraction Layer)

A software-hardware interface that enables the FHE Inference Engine to run efficiently on both off-the-shelf accelerators (e.g. GPUs) and designated FHE hardware.