April 2023 Summaries

Large language models (LLMs) represent a significant advancement in AI, capable of transforming various sectors by processing contextual data, such as personal or sensitive information. However, this raises privacy concerns, leading to bans in countries like Italy. Fully Homomorphic Encryption (FHE) offers a solution by allowing encrypted data processing without decryption, ensuring privacy while leveraging AI capabilities. Despite past limitations in FHE's speed and practicality, advancements in cryptographic techniques and hardware acceleration promise significant improvements, potentially making end-to-end encrypted LLMs viable within five years. This development could render privacy concerns obsolete, as data protection would be inherently guaranteed.

TFHE-rs v0.2.0 introduces significant enhancements such as support for large homomorphic unsigned integers up to 256 bits, a high-level API to simplify builds, and an upgraded parallelized Programmable Bootstrapping (PBS) to improve latency. The update allows operations on large precision integers, including additions, subtractions, multiplications, and bitwise functions, using Radix or Residue Number Systems for representation. The high-level API, inspired by Rust, facilitates operations on homomorphic unsigned integers by wrapping arithmetic operators and mimicking clear operations, while a C API aids integration into existing applications. Additional features include a choice of encryption keys, a reduction in public encryption key size, and the ability to compute PBS over 128-bit ciphertexts, with a parallel PBS implementation based on recent research. Future updates are expected to include faster arithmetic operators and further cryptographic features, with ongoing improvements in performance.

Zama has released updates to its suite of homomorphic encryption products, including TFHE-rs v0.2.0, Concrete v1.0.0, and Concrete ML v1.0.0. The TFHE-rs v0.2.0 update enhances the Rust-based library by supporting large homomorphic unsigned integers up to 256 bits, adding a high-level API, and incorporating a parallelized version of Programmable Bootstrapping to reduce latency. Concrete v1.0.0, now fully open-source, offers the capability to evaluate any function without approximation, supports multiple hardware backends, and includes automatic cryptographic parameter optimization. Meanwhile, Concrete ML v1.0.0 focuses on improved inference performance, user-friendly error reporting, and streamlined model deployment in cloud environments, making these encryption tools more accessible and efficient.

Concrete v1.0.0 is a newly released compiler designed for transforming Python programs into their Fully Homomorphic Encryption (FHE) equivalents using the TFHE scheme, particularly the Zama variant. It facilitates the development of applications that process encrypted inputs to produce encrypted outputs, easing the challenge of converting regular programming operations into encrypted forms. Notably, it supports operations through a Directed Acyclic Graph, optimizing cryptographic parameters automatically and offering parallelization capabilities to enhance performance. The update renames the Concrete Numpy package to Concrete Python and transitions the Concrete Rust library to the TFHE-rs library. Future improvements aim to extend language support beyond Python, reduce key and ciphertext sizes, and increase precision. The platform is also evolving to include multiple hardware backends, with current support for CPU and upcoming support for GPU, making it adaptable for different target architectures.

Concrete ML v1.0.0 introduces significant advancements such as a stable API, enhanced inference performance, and improved user-friendly error reporting, making it easier to deploy models in cloud environments. The update aligns the Concrete ML API with Concrete, ensuring forward compatibility but possibly requiring code updates, for which a transition guide is provided. Focusing on Fully Homomorphic Encryption (FHE), the release includes better tools for designing FHE-compatible models and a new roundPBS feature that dramatically improves latency without losing accuracy by optimizing the application of activation functions. This version also introduces structured pruning in neural networks to accelerate inference times in FHE applications. Easy deployment is emphasized, with tutorials available for setting up FHE executions on AWS, and new resources such as a Hugging Face application for encrypted image filtering demonstrate the capabilities of Concrete ML. Additionally, technical articles provide insights into privacy-preserving tree-based inference and neural networks for encrypted inference, contributing to a deeper understanding of ML with FHE.