October 2024 Summaries

A group of computer science students from the National University of Singapore participated in the 2024 TikTok TechJam, where they developed a privacy-preserving ad-serving system using Fully Homomorphic Encryption (FHE) during the hackathon. Their project, AnonymousAds, aimed to provide personalized ad targeting while ensuring user privacy by encrypting user data on the client side before sending it to the server for processing. The server, using Zama's Concrete ML, processed the encrypted data to predict user interest categories without accessing the actual data. This approach allowed the team to maintain user privacy, as the server never accessed plain text data, and employed noise injection techniques to further obscure user interests. The project involved various components, including a client-side search engine interface, server-side encrypted data processing, and the use of a neural network regressor model, all managed within a containerized environment using Docker. The students' innovative approach to solving the privacy concerns in online advertising led to their project winning the competition, showcasing the transformative potential of FHE in the digital advertising industry.

TFHE-rs is a library designed for developers and researchers seeking control over TFHE functionalities without delving into low-level implementations, aiming for stability, simplicity, high performance, and readiness for production. The tutorial by Zama team member Arthur Meyre provides guidance on performing array and tensor operations on encrypted data using TFHE-rs. The initiative encourages community engagement through various means, such as starring the GitHub repository, reviewing documentation, participating in support channels, contributing to the progress of Fully Homomorphic Encryption (FHE) through Zama's Bounty and Grant Programs, and engaging in a developer survey.

Concrete ML is a set of tools designed to facilitate the use of Fully Homomorphic Encryption (FHE) by developers, enabling the automatic conversion of machine learning models into their homomorphic counterparts. In a tutorial, Zama team member Celia Kherfallah demonstrates how to create encrypted DNA testing using FHE and Concrete ML. The initiative encourages community engagement through various channels, such as starring the Concrete ML GitHub repository, reviewing documentation, seeking support on the Community Forum, and contributing to the advancement of FHE through the Zama Bounty Program and a developer survey.

Concrete is an open-source fully homomorphic encryption (FHE) compiler designed to simplify the development of FHE programs by integrating with the TFHE Compiler based on LLVM. In a tutorial, Zama team member Ayoub Benaissa demonstrates how to combine Concrete with TFHE-rs computation using Concrete's interoperability feature. The framework encourages community engagement through various channels, such as endorsing their work on GitHub, reviewing documentation, seeking support, and participating in initiatives like the Zama Bounty Program and a developer survey to further advance the FHE space.

Zama has announced the release of several product updates, including TFHE-rs v0.8, Concrete v2.8, and Concrete ML v1.7, with an upcoming release of fhEVM v0.6 in November. The latest version of TFHE-rs introduces array types and improves multi-GPU computing, significantly accelerating arithmetic operations on GPUs. Concrete v2.8 enhances interoperability with TFHE-rs and provides automatic module tracing for easier module compilation fine-tuning, along with documentation improvements and GPU runtime optimizations. Concrete ML v1.7 enables fine-tuning of LLMs and neural networks on encrypted data, enhances performance with GPU acceleration, and supports Python 3.11 and PyTorch 2, showcasing privacy-preserving machine learning use cases like predicting ancestry from encrypted DNA. The upcoming fhEVM v0.6 will feature a new Coprocessor that allows confidential smart contracts to be deployed directly on Ethereum, leveraging homomorphic encryption for enhanced on-chain privacy, security, and scalability with full compatibility with the Ethereum ecosystem.

TFHE-rs v0.8 introduces significant enhancements with the addition of array types and improved multi-GPU computing capabilities, enabling more efficient operations on encrypted data. This update allows developers to work seamlessly with vectors and tensors, facilitating operations like element-wise arithmetic and bitwise functions on encrypted arrays. The improved multi-GPU support eliminates previous limitations, enabling the use of all Nvidia GPUs, including those connected via PCIe, and optimizing inter-GPU communication, which significantly speeds up computation, particularly for arithmetic operations like 64-bit integer multiplication. The release also includes features such as post-homomorphic computation ciphertext compression on GPUs to reduce memory usage, new GPU-based homomorphic operations, and improved bootstrapping performance. Additionally, enhancements in CPU operations have been made, along with the introduction of new operations like parity detection and encrypted random FheBool generation. This version focuses on improving code stability and introduces various features, with future releases set to bring new data types and continued performance enhancements.

Concrete v2.8 introduces significant enhancements like interoperability between TFHE-rs and Concrete, allowing developers to leverage the strengths of both platforms in handling different integer sizes within Fully Homomorphic Encryption (FHE) applications. The update also simplifies module compilation with automatic module tracing, reducing the complexity of defining composition rules and input sets by aligning them with the natural flow of Python code. The release includes optimizations, especially in the Concrete GPU runtime, which enhance the speed of FHE evaluations without requiring changes to the API. New tutorials illustrate practical applications such as private information retrieval and floating point management, while ongoing improvements and community engagement are encouraged through various support channels and the Zama Bounty Program.

Concrete ML v1.7 introduces significant advancements, including the ability to fine-tune large language models (LLMs) and neural networks on encrypted data, thus enhancing privacy-preserving machine learning applications. This version also supports GPU acceleration, significantly boosting performance for encrypted inference on large models like GPT and ResNet, particularly when using GPUs such as NVIDIA H100. Additionally, a new Hugging Face space demonstrates predicting ancestry from encrypted DNA, highlighting the potential of privacy-centric genetic analysis. Concrete ML v1.7 is compatible with Python 3.11 and PyTorch 2, further broadening its usability. The update ensures that sensitive computations are securely outsourced, while user-side fine-tuning maintains data privacy, making it a notable step forward in secure machine learning practices.