January 2023 Summaries

On January 17, 2023, the Zama team presented Concrete ML during a Google Tech Talk, highlighting their open-source library that facilitates the conversion of Machine Learning models into Fully Homomorphic Encryption (FHE) counterparts to enhance data privacy. Concrete ML aims to enable zero-trust interactions between clients and service providers, allowing ML models to be deployed on untrusted servers without compromising user data privacy. Google Tech Talks is a platform designed to share diverse technical discussions, featuring expert insights across various fields, and the views expressed are those of the speakers and not necessarily of Google, Inc.

TFHE-rs is a new Rust implementation of TFHE, designed for developers and researchers working with encrypted data to provide an enhanced, flexible framework for boolean and small integer arithmetic operations. It aims to serve as the reference implementation for TFHE, offering up-to-date security features and easy integration for application developers and compiler developers. The library includes powerful features like a low-level cryptographic library, a boolean API for chaining infinite boolean gates, and a short integer API for unbounded integer arithmetic with up to 8 bits of message space. TFHE-rs also supports public key encryption, ciphertext compression, and comes with a full Rust and C API, along with a client-side WASM API. Distinct from Zama's previous FHE products under the "Concrete" name, TFHE-rs provides full control over FHE circuit execution, while the refactored Concrete is now focused on compiler use for optimal circuit performance. Future updates will include APIs for large-integer and floating-point arithmetic, along with performance enhancements. The library is benchmarked as the fastest public implementation of TFHE, supporting AVX512 acceleration for efficient operations.

Zama has released TFHE-rs, the fastest public implementation of the TFHE scheme, along with updates to its Concrete ML and Concrete Numpy libraries, as part of its ongoing effort to make fully homomorphic encryption more accessible, efficient, and user-friendly. TFHE-rs is a pure Rust implementation designed for boolean and small integer operations on encrypted data, serving as the new reference implementation with advanced security parameters and features. The updated Concrete Numpy v0.9 includes performance improvements, 16-bit integer table lookups, and additional new features, while Concrete ML v0.6 introduces fast, high-precision linear models and support for 16-bit accumulators for both built-in and custom neural networks.

Zama has released a new version of Concrete-ML, incorporating features like fast, high-precision linear models and support for 16-bit accumulators, enhancing the performance of both built-in and custom neural networks. The update allows encrypted values to significantly boost the accuracy of neural networks, particularly in complex computer vision tasks, as demonstrated by tutorials and use-case examples on datasets like CIFAR10 and CIFAR100. The release optimizes Concrete-ML for performing only linear computations on encrypted data, achieving high precision and low latency. Concrete-ML also introduces better simulation and debugging features, allowing users to simulate the performance of models using Fully Homomorphic Encryption (FHE) without running time-consuming computations. Additionally, a live demo on Hugging Face Space illustrates the real-time application of FHE in analyzing encrypted texts for sentiment inference, supported by detailed documentation and community engagement channels.

Zama's release of Concrete-Numpy v0.9 brings significant updates, including improvements in performance with the updated Concrete-Compiler v0.23, and expanded support for table lookups on integers up to 16 bits. The update introduces a new "Direct Definition" compilation method, which simplifies the process for certain use cases and supports all operations of traditional compilation. Enhancements also include the ability to handle signed inputs without additional table lookups, simulate local error probability in virtual circuits, and provide a global error probability option. The release features tagging for easier debugging of large circuits and improved error messages that pinpoint the exact location of issues. New operations like encrypted multiplication and maxpool, better support for numpy functions like np.expand_dims and np.transpose, and various bug fixes further enhance the utility and robustness of Concrete-Numpy for encrypted computation.

The text discusses the complexity and intricacies of Fully Homomorphic Encryption (FHE), contrasting it with the more straightforward RSA encryption scheme. It emphasizes that while RSA is commonly used as a beginner's introduction to cryptography, it is often oversimplified and rarely used in its textbook form due to security vulnerabilities. Instead, FHE should be viewed as a multi-party computation (MPC) protocol rather than a simple encryption scheme, involving multiple entities such as Alice, Bob, Charlie, and an adversary, Eve. Each entity plays a role in the encryption, decryption, and computation process, and the protocol must accommodate potentially dishonest participants. The text highlights the need for additional cryptographic constructs like zero-knowledge proofs and verifiable computation to ensure the integrity and security of FHE. It also discusses strategies such as using multiple parties to avoid single points of failure and underscores the importance of viewing FHE within a broader cryptographic context to effectively manage adversarial threats.