Andrei Stoian May 2024

Concrete ML
Privacy Preserving Machine Learning
with Fully Homomorphic Encryption
 Zama is a cryptography company providing open
About Zama

source homomorphic encryption solutions for

blockchain and AI.
 Encrypted data E[x] is
sent to the evaluator, who
cannot decrypt it

Fully
Homomorphic
Encryption
About Zama

Evaluator computes f(E[x])

(FHE) enables Encryptor encrypts data
x with its secret key
blindly, producing an
encrypted response E[f(x)].
processing
data blindly Encrypted response
E[f(x)] is sent to the
encryptor, who can
decrypt it
 E[x] + E[y] = E[x + y]

FHE enables E[x] < E[y] = E[x < y]

About Zama

encrypted data
processing More generally:
f(E[x], …, E[y]) = E[f(x, …, y)]
 Zama FHE performance
(Million FHE gates / $)

500
FHE is getting
exponentially faster
FHE smart contracts are doable today, with a
About Zama

throughput of ~5 tps. FHE hardware will enable

1,000+ tps at a fraction of the cost.

150

Mass
adoption
40
11
3

2019 2021 2023 2025 2027

hardware
 Our products enable secrecy-preserving,
regulatory-compliant applications

Detection

Auctions
Scoring

Token
Credit

Fraud

Swap

Blind
...

...

Collaboration
About Zama

Verification
Concrete ML Confidential Blockchain

Biometric
Data
fhEVM

...
Open-source privacy preserving
Machine Learning framework FHE smart contract protocol

Concrete & TFHE-rs

Open Source compiler and library to generate TFHE circuits and use TFHE operations

CPU GPU FPGA Other

Zama's Products Third Party Applications Hardware
 Concrete ML
Concrete ML
 Concrete ML enables data scientists
Concrete ML

to work with sensitive data while

protecting their models
 When should you consider
confidential ML?
Concrete ML

Working with private or confidential data

Working with regulated data subject to
data residency
You care about data privacy
The model has sensitive IP
 What can confidential ML
do for you?

Train a model without seeing the

Concrete ML

confidential training data

Run inference without seeing the
confidential data or the inference results
Protect the intellectual property of your
model
 How does Concrete ML
achieve confidentiality?

Data and results are always encrypted with

Concrete ML

post-quantum encryption FHE allows the

processing of encrypted data without
decrypting it
No need to trust a hardware manufacturer
No need to manage multiple, independent
trusted entities
 Use-cases

Private Inference
Concrete ML

Secure Collaboration

Private Training

Private LLMs
 Concrete ML: Powerful
features are available
out of the box
Supports many types of models
Linear models, XGBoost, trees, neural networks...

Unlimited number of layers

Concrete ML

Unlimited depth without FHE noise accumulation

Support for any activation function

Use any non-linear activation without approximation

Automatic selection of FHE parameters

Model-derived FHE quantization and parametrization

Exact results, just like in plaintext

FHE output is exactly the same as plaintext one

Confidential (& collaborative) Training

Train in FHE or combine with Federated Learning
 Concrete ML
An open-source library for data scientists, for PPML
Built for data scientists on top of scikit-learn and torch
No need to know cryptography
Concrete ML

Large Language Models

Linear Models Tree-Based Models Neural Networks

Dataframe
 Easy to use APIs
Zama Products

No need to know
cryptography!
 Easy to use APIs

Custom
Neural
Networks
Tensorflow
Concrete ML

Torch
ONNX
 Easy to use APIs

Image Filtering
A Hugging Face space to apply filters over images homomorphically using the
Concrete-ML Presentation

developer-friendly tools in Concrete-Python and Concrete-ML.

Illustration here
Illustration here

huggingface.co/spaces/zama-fhe/encrypted_image_filtering
 Easy to use APIs
Image
Filtering
Concrete-ML Presentation

huggingface.co/spaces/zama-fhe/encrypted_image_filtering
 Easy to use APIs

Sentiment
Concrete-ML Presentation

Analysis

huggingface.co/spaces/zama-fhe/encrypted_sentiment_analysis
 Easy to use APIs
Sentiment
Analysis
Concrete-ML Presentation

huggingface.co/spaces/zama-fhe/encrypted_sentiment_analysis
 Easy to use APIs
Anonymized ChatGPT queries
Concrete-ML Presentation

https://huggingface.co/spaces/zama-fhe/encrypted-anonymization
 Easy to use APIs
Anonymized ChatGPT queries
Concrete-ML Presentation

https://huggingface.co/spaces/zama-fhe/encrypted-anonymization
 Easy to use APIs

Private DNA Ancestry

Concrete-ML Presentation

Client Side Server Side Client Side

Predictor
12% West African
34% East African
Smoother 5% East Asian
 Easy to use APIs

Private DNA Ancestry

Concrete-ML Presentation

Model definition

Model training
 Easy to use APIs
Style transfer
Concrete-ML Presentation
 Easy to use APIs
Style transfer
Concrete-ML Presentation

Your solution here!

YouYou

Join the Zama bounty program

github.com/zama-ai/bounty-and-grant-program
 zama.ai
Contact
and Links Github

Community links
Concrete ML