April 2022 Summaries

Concrete Numpy, an open-source toolset designed to facilitate the use of fully homomorphic encryption (FHE) for data scientists, has released a new version, v0.5, which introduces several enhancements and changes. This version supports additional numpy operators, including sum, concatenate, 2D convolution, matmul, and transpose, while also offering increased bit precision from 7 bits to 8 bits. An important shift is the removal of some higher-level machine learning features, which are now part of a new package called concrete-ml. The update also includes enhanced support for loop parallelism and a refined workflow featuring explicit encrypt, decrypt, and run functionalities, replacing the previous run command with a new encrypt_run_decrypt process. Built on the Concrete library's foundation of efficiency, usability, and simplicity, this version aims to improve the experience of training and deploying machine learning models using FHE, with a separate API for key generation, encryption, decryption, and execution.

Zama has introduced its new product, Concrete ML version 0.2, as well as updates for Concrete Numpy version 0.5 and Concrete Core version 1.0-beta, enhancing its product line aimed at making fully homomorphic encryption (FHE) more accessible and efficient. Concrete ML is designed to simplify the process of compiling machine learning models into their FHE equivalents and integrates with popular frameworks like Scikit-Learn and PyTorch. Meanwhile, the updated Concrete Numpy offers improved operators, precision, and performance, and Concrete Core's latest version features new public engines, entities, and two companion libraries named Concrete-core-fixture and Concrete-core-test. These developments underscore Zama's commitment to advancing FHE technology through user-friendly and high-performance solutions.

Concrete-core V1.0.0-beta has been released, marking a significant step toward the official 1.0.0 release, and introduces several new features and enhancements aimed at expanding cryptographic operations and simplifying testing and integration. This beta version includes new public engines and entities in the Core backend, such as the packing keyswitch, the encryption of GGSW ciphertexts, and engines for trivial encryption and decryption. Additionally, two new companion libraries, Concrete-core-fixture and Concrete-core-test, are introduced to facilitate the systematic testing of engines across different backends by using prototype generation, synthesis, and result verification processes. These additions enhance the ability to test and compare new implementations and hardware accelerations, making it easier to integrate new features into the Concrete-core framework. The release aims to streamline the development process for hardware acceleration in Fully Homomorphic Encryption (FHE) and invites developers to explore further through tutorials and contributions.

Concrete ML has been released as a public alpha, built on top of Concrete Numpy, to enable data scientists with no cryptography knowledge to convert classical machine learning models into their Fully Homomorphic Encryption (FHE) equivalents. The release aims to simplify the adoption process for users of popular ML frameworks by providing user-friendly APIs that allow the seamless integration of FHE capabilities, as demonstrated through examples with linear models and tree-based classifiers. While tree models perform excellently with encrypted data due to Zama's Programmable Bootstrapping, linear models and neural networks currently face performance challenges, which are expected to improve with future updates and enhancements in quantization and precision. The package also supports user-provided torch models, utilizing an ONNX conversion pipeline to facilitate the use of a wide range of operators while maintaining focus on feature-completeness. As efforts continue to enhance the performance of various models under FHE constraints, Concrete ML presents a promising solution for private computations in machine learning.

In 2015, the author met Nigel, whose practical approach to technology influenced the development of an open-source protocol for private analytics on mobile devices, preceding the widespread emphasis on privacy and GDPR. As co-founder of Zama, the author, alongside Pascal, sought Nigel's expertise, leading to his involvement as an academic advisor and now his official role as Chief Academic Officer (CAO). Nigel's role focuses on ensuring Zama's open-source technology is utilized in academia, fostering advancements in homomorphic encryption research, and integrating breakthroughs into the open-source codebase for developers. Nigel Smart, a renowned cryptography researcher and entrepreneur, has contributed significantly to fields like FHE and MPC and has founded several successful companies. Zama is expanding, aiming to hire up to 100 new employees to advance internet security, with updates available via their newsletter.