January 2025 Summaries

TFHE-rs v0.11 introduces significant updates and enhancements, including improved performance for Zero Knowledge Proofs and a new module for encrypted strings accessible through the FheAsciiString type in the High-Level API. The release features a shift to a tweaked uniform (TUniform) noise distribution for default cryptographic parameters, enhancing boundedness compared to the previous Gaussian distribution. GPU performance has been markedly improved, with 64-bit addition operations now 30% faster, and the ability to conduct computations on arrays of ciphertexts more efficiently on GPU. A highlight is the FHE strings module, a contribution from the Zama Bounty Program, allowing operations akin to Rust’s str type. Additionally, the release includes support for n-dimensional arrays on GPU, facilitating operations like vector comparison and searching within encrypted data. Other enhancements include new GPU operations, faster CPU operations, and a focus on code stability. The update aims to enhance the capabilities and efficiency of homomorphic encryption computations, with further developments anticipated in future releases.

fhEVM v0.6 introduces significant advancements in its type support, input mechanism, and configurability, enhancing its integration with encrypted data systems and overall usability. The update includes support for larger encrypted types, like euint128 and euint256 for higher-precision integers, and ebytes64, ebytes128, and ebytes256 for larger byte arrays. It also features an Input Mechanism with Proof to ensure ciphertext authenticity and integrity, preventing unauthorized manipulation. Enhanced configurability is achieved through the [.c-inline-code]TFHE.setFHEVM()[.c-inline-code] function, allowing users to specify fhEVM configurations on networks easily. A notable innovation is the fhEVM Coprocessor, which enables the deployment of confidential smart contracts on any EVM chain using Fully Homomorphic Encryption (FHE), without altering the blockchain protocols. Initially available on the Sepolia testnet, the fhEVM Coprocessor paves the way for more secure and versatile blockchain applications, with support for additional chains forthcoming.

Concrete ML v1.8 introduces significant advancements in privacy-preserving fine-tuning for Large Language Models (LLMs) by enhancing the speed and usability of hybrid fine-tuning through an optimized Fully Homomorphic Encryption (FHE) backend and a new Low Rank Approximation API. This version now supports Python 3.12 and features a streamlined fine-tuning API inspired by HuggingFace PEFT LoraTrainer, allowing efficient use of GPUs for model evaluation and development acceleration. The FHE backend speeds up computation by leveraging GPU resources, implementing encrypted matrix - clear matrix multiplication, and achieving efficient compression, which results in encrypted data sizes only four times larger than non-encrypted data. Fine-tuning a LLAMA 8B model on 100,000 tokens takes around 70 hours at an estimated cost of $500 using a decentralized network of 100 consumer-grade GPUs, with expectations of further optimizations to reduce costs and latency. Concrete ML v1.8 thus facilitates secure, scalable, and decentralized AI solutions, with future updates promising continued improvements.

The Concrete team is enhancing interoperability between TFHE-rs and Concrete by expanding support for Python 3.12 and improving functionality on both frontend and backend. Concrete v2.9 builds on the previous version by extending support for signed integers and fully supporting the serialization and deserialization of tensors, thereby enabling more complex computations, such as linear machine learning models. These updates include performance optimizations and bug fixes, particularly in GPU kernels now hosted on the TFHE-rs GitHub repository, leading to improved performance in large FHE evaluations without requiring API changes. The release is available for Python versions 3.8 to 3.12 on Linux and Mac OS platforms and encourages community feedback to further improve the accessibility and efficiency of fully homomorphic encryption (FHE).

Zama has released new versions of its products, including TFHE-rs v0.11, Concrete v2.9, Concrete ML v1.8, and fhEVM v0.6, all designed to enhance the accessibility, ease, and speed of homomorphic encryption. TFHE-rs v0.11 features improved performance for Zero Knowledge Proofs, a new module for encrypted strings, and better GPU performance, while adopting a TUniform noise distribution. Concrete v2.9 enhances interoperability with TFHE-rs and supports Python 3.12, applicable in both development and production environments. Concrete ML v1.8 advances privacy-preserving fine-tuning for Large Language Models by optimizing the FHE backend and introducing a Low Rank Approximation API, along with Python 3.12 compatibility. The updated fhEVM v0.6 broadens type support, improves input mechanisms, and introduces the fhEVM Coprocessor, enabling developers to create confidential smart contracts on any EVM chain, thus boosting the flexibility and integration of encrypted data systems.