March 2025 Summaries

Duality Technologies v4.3 offers a robust solution for secure data collaboration in AI, particularly in sensitive sectors like healthcare and finance, by integrating NVIDIA FLARE, Google Cloud Confidential Space, and proprietary privacy-enhancing technologies into a federated learning platform. This platform simplifies deployment across organizations and ensures secure data handling and preprocessing, project management, and enforcement of privacy policies, including differential privacy. By utilizing Trusted Execution Environments (TEEs) for secure aggregation, Duality protects against information leakage during federated learning, allowing organizations to collaboratively train AI models without exposing sensitive data. The platform's integration of NVIDIA FLARE and Google Cloud Confidential Space enhances the security and efficiency of federated learning workflows, supporting various real-world applications like early cancer detection and fraud detection, while maintaining regulatory compliance and data protection.

Since the inception of Fully Homomorphic Encryption (FHE) by Craig Gentry in 2009, multiple generations of FHE schemes have emerged, each improving the efficiency of encrypted computations and facilitating various real-world applications. Initial practical schemes like BFV and BGV were tailored for shallow arithmetic operations due to costly bootstrapping, with applications such as Private Information Retrieval and Private Set Intersection. Hybrid systems like FHEW and TFHE introduced efficient bootstrapping after each operation, enabling privacy-preserving applications but facing scalability issues. The CKKS scheme offered efficient computations over encrypted real or complex vectors, with significant throughput advantages, particularly beneficial for machine learning tasks, though previously lacking in general functional bootstrapping. The latest breakthrough, developed by Duality Technologies in collaboration with Andrey Kim, introduces a vectorized hybrid cryptosystem combining BFV and CKKS to perform batched functional bootstrapping with a throughput significantly higher than previous methods. This advancement enhances the applicability of Secure AI applications on large datasets, improving performance by threefold, and will be integrated into the OpenFHE library, promising substantial impacts on fields like neural network inference, encrypted SQL, and Web3 applications.