Meet Concrete ML: An Open-Source FHE-Based Toolkit That Helps in Preserving Privacy And Enables Secure Machine Learning

Artificial Intelligence and Machine Learning have shown tremendous productivity rise in the past few years. ML is all about having good quality data by maintaining all means of privacy and confidentiality. It is very important to bridge the gap between privacy and utilizing the advantages of Machine Learning in order to solve problems. In today’s data-driven days, protecting one’s privacy has become very difficult. With Machine Learning becoming so prevalent nowadays, the implications must be taken care of, and safeguarding clients’ information is necessary. New advancements like Fully Homomorphic Encryption (FHE) have successfully protected user information and maintained confidentiality.

Machine Learning researchers at Zama have introduced an open-source library called Concrete-ML which allows the smooth conversion of ML models into their FHE counterparts. They have recently presented Concrete ML during a Google Tech Talk. Whenever some of the data belonging to the user are sent to the cloud, Homomorphic encryption schemes protect that data. The operations and all the actions take place over encrypted data by considering data safety. Fully Homomorphic Encryption can be explained with the help of an example. Say a doctor wishes to evaluate descriptive analytics on patients suffering from heart issues in a particular city. The internal team of the hospitals in that city that safely stores the patient data in their databases might be unable to reveal the data because of privacy concerns. That is where FHE encrypts the sensitive data so that the data is safe as well as computing.

Concrete ML is an open-source toolkit that has been developed on top of The Concrete Framework. It helps researchers and data scientists automatically convert Machine Learning models into their identical homomorphic units. The key feature of Concrete ML is its ability to turn ML models into their FHE equivalent without necessarily having any previous knowledge about cryptography. With Concrete ML, users are able to have zero-trust conversations with different service providers without hampering ML models from getting deployed. The privacy of the data and the user is maintained, and ML models are put into production on even untrusted servers.

FHE, an encryption strategy that permits direct computing on encrypted data, can be used to develop applications with unique features. FHE doesn’t require the need for decryption. Concrete ML uses some popular Application User Interfaces (API) from Scikit-learn and PyTorch. The Concrete ML model has been designed in the following way –

1. Training of the model – The model is trained on some unencrypted data using the Scikit-learn library. Concrete ML only uses integers during the inference, as FHE only works over integers.
2. Conversion and compilation – In this step, the model is converted into a Concrete-Numpy program, followed by the compilation of the quantized model into an FHE equivalent.
3. Inference – The inference is conducted on the encrypted data. During the deployment of the model on the server, the data is encrypted by the client, followed by secure processing by the server and decryption by the client.

Concrete ML is a great development in using Machine learning with complete privacy and trust. While currently, the only limitation Concrete ML has is that it can only run within the supported precision of 16-bit integers, it still sounds promising for privacy preservation.

Check out the Github Link. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 14k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.