July 2024 Summaries

Concrete is an open-source compiler designed to simplify the use of fully homomorphic encryption (FHE), which enables computations on encrypted data without decryption. The framework includes a TFHE Compiler based on LLVM, facilitating the development of FHE programs. In a tutorial, Antoniu Pop from the Zama team demonstrates implementing GPU acceleration in FHE using Concrete. Users are encouraged to support the project by starring its GitHub repository, reviewing the documentation, engaging in community channels, and participating in the Zama Bounty Program to further the field of FHE.

TFHE-rs is a library designed for developers and researchers seeking comprehensive control over their use of the Fully Homomorphic Encryption (TFHE) scheme without dealing with complex low-level implementations. Its aim is to provide a stable, simple, and high-performance library equipped with advanced TFHE features, making it suitable for production environments. The tutorial by Zama team member Arthur Meyre focuses on using TFHE-rs to compress ciphertexts after FHE computations. Users are encouraged to engage with the TFHE-rs GitHub repository, explore its documentation, participate in community channels for support, and contribute to the advancement of the FHE field through the Zama Bounty Program.

Concrete ML is a set of tools designed to facilitate the use of Fully Homomorphic Encryption (FHE) in machine learning by enabling developers to automatically convert machine learning models into their homomorphic equivalents. A tutorial by Zama team member Jordan Frery focuses on improving the latency of larger neural networks within Concrete ML. Users are encouraged to support the project by starring the Concrete ML GitHub repository, reviewing documentation, engaging on community channels, and participating in the Zama Bounty Program to further advance the field of FHE.

Zama has made significant advancements in speeding up Fully Homomorphic Encryption (FHE) for machine learning by developing Concrete ML, which has surpassed previous performance benchmarks. Initially, Zama used an ONNX-based compiler prototype to convert neural networks into FHE-friendly models, demonstrating the potential of the TFHE cryptographic scheme for handling deep neural networks. However, the complexity of integrating machine learning and cryptography in a single compiler led to a strategic pivot towards a more generic framework using MLIR, which allowed for better support of hardware accelerators and improved system efficiency. This transition enabled a separation of machine learning tasks from cryptographic ones, focusing on cryptographic security while letting users manage machine learning-specific choices. The new approach, utilizing the Concrete ML and Concrete Python frontends, offers flexible support for various backends and has achieved faster execution times for neural network models such as NN-20 and NN-50 compared to earlier methods. Zama's continued efforts in improving the TFHE-rs library, leveraging the MLIR framework, and enhancing quantization techniques have contributed to these improvements, making FHE more accessible and efficient for real-world applications.

Season 5 of the Zama Bounty Program challenged participants to develop a machine learning system using Fully Homomorphic Encryption (FHE) to determine ancestry from encrypted DNA data, highlighting the need to protect sensitive genetic information. Two developers, Alephzerox and Soptq, shared first prize with distinct yet effective solutions. Soptq employed a logistic regression-based method, segmenting chromosome data into windows to predict ancestry, achieving 96% accuracy with a latency of around 300 seconds for encrypted genomes. Alephzerox, on the other hand, implemented a similarity search approach using a reference panel of pure-blooded genomes to determine ancestry, also reaching 96% accuracy with an accuracy increase proportional to the number of reference genomes. While both methods achieved similar accuracy, Soptq's machine learning approach led to lower inference latency in processing encrypted data, showcasing the potential of FHE in protecting DNA information.

Concrete ML v1.6 introduces significant updates that enhance performance and usability, including reduced latency for large neural networks, support for pre-trained tree-based models, and improved collaborative computation through DataFrame schemas and logistic regression deployment. The update supports importing pre-trained tree models with the [.c-inline-code]from_sklearn[.c-inline-code] function while maintaining accuracy on encrypted data. Notable latency improvements are showcased in two notebooks, demonstrating a 20-layer deep MLP model with a significant reduction to 1-second latency on encrypted data and a ResNet18 model showing a 4x improvement over previous results. Additionally, deployment enhancements allow logistic regression training to be easily deployed as a client-server service, with options for parametrization and cloud deployment. DataFrame schemas now enable users to control schema details, facilitating compatibility across different users' encrypted data. Upcoming GPU support promises further advancements, and users are encouraged to engage with the project through various community channels and participate in the Zama Bounty Program.

Zama has released updated versions of its homomorphic encryption products, including TFHE-rs v0.7, Concrete v2.7, Concrete ML v1.6, and fhEVM v0.5, advancing the accessibility, efficiency, and security of encrypted computations. TFHE-rs v0.7 introduces a significant reduction in ciphertext size and enhanced computational performance through multi-GPU support. Concrete v2.7 adds GPU acceleration and extends function composition capabilities, alongside new Python frontend features. Concrete ML v1.6 improves latency for large neural networks, facilitates collaborative computation, and supports pre-trained tree-based models, with upcoming GPU support. fhEVM v0.5 focuses on security improvements and introduces a stable Solidity API for future compatibility across various chains, enhancing how encrypted data is handled in applications.

fhEVM v0.5 introduces significant advancements aimed at enhancing the security and efficiency of encrypted data handling in applications, notably through mechanisms such as packed input, an enhanced Access Control List (ACL), and updated decryption and re-encryption processes. The update allows multiple plaintext values to be packed into a single ciphertext, facilitating future developments like proven inputs, while the new ACL system provides robust management of ciphertext permissions, ensuring that only authorized entities can manipulate encrypted data. Security is further bolstered by the use of an off-chain service for re-encryption, optimizing cryptographic operations and allowing for batch decryption of multiple types. The release finalizes the Solidity API, ensuring compatibility across future versions and various blockchain platforms, and introduces a new permissions model and deprecates older functions in favor of updated methods. Additionally, fhEVM v0.5 is designed to accommodate future enhancements such as zero-knowledge proofs and new types, thus promoting long-term stability and broad usability for developing confidential smart contracts.

Concrete v2.7 introduces significant advancements, including the first wheel enabling GPU acceleration for Fully Homomorphic Encryption (FHE), which facilitates faster computations for certain workloads, such as those involving heavy linear algebra. This release also enhances function composition by optimizing modules through a partition-based approach, allowing different cryptographic parameters for various parts of a module, which results in improved performance. Users can leverage GPU acceleration by installing the GPU-specific wheel via a dedicated PyPI repository and setting the use_gpu option during compilation, although the performance gains depend on the specific hardware and workload involved. Additionally, this version allows for more precise dependency specification between module functions, enabling tighter cryptographic parameters and faster execution. Alongside these major updates, several minor improvements have been made, all of which improve the accessibility and efficiency of FHE, as detailed in the GitHub release notes and documentation.

TFHE-rs v0.7 introduces significant advancements in fully homomorphic encryption (FHE) by introducing compressed ciphertexts and support for multi-GPU architectures. With this update, ciphertexts resulting from homomorphic computations can be compressed, reducing their size by up to 1,900 times, a crucial improvement given the typically large size of encrypted data. The release also leverages multi-GPU setups to enhance computational performance without requiring users to modify their code, although it is limited to GPUs with peer access via NVLink. Additionally, the update includes new cryptographic parameter sets, vector and array operations, improved zero-knowledge proofs, and optimized keyswitch operations on GPUs, collectively enhancing the efficiency and functionality of TFHE-rs. These improvements not only reduce the latency of programmable bootstrapping but also expand the capabilities of the tool, promising further enhancements in future updates.