At a Glance
- Researchers at NYU Tandon have developed Orion, a deep learning framework that uses fully homomorphic encryption to allow AI models to compute on encrypted data without compromising privacy or security.
- Unlike traditional encryption, which only protects data during storage or transfer, FHE enables the processing of encrypted data, and Orion automates the conversion of PyTorch models into efficient FHE-compatible programs.
- Orion demonstrated a 2.38-times performance improvement over existing methods on the ResNet-20 model and successfully handled high-resolution object detection using the much larger YOLO-v1 architecture with 139 million parameters.
- This advancement is especially relevant for industries like healthcare and finance. It enables privacy-preserving AI applications that process sensitive information without revealing raw data to the computing system.
- By open-sourcing Orion, the researchers aim to make fully homomorphic encryption more accessible, potentially setting a new standard for secure AI development across various sectors.
In the digital age, where data privacy is a significant concern, a breakthrough in artificial intelligence (AI) could significantly enhance the security of sensitive information. Researchers at the NYU Tandon School of Engineering have introduced Orion, a novel framework that integrates fully homomorphic encryption (FHE) with deep learning. This development allows AI models to perform computations on encrypted data without decrypting it, maintaining privacy while enabling advanced data processing.
FHE is a form of encryption that allows data to remain encrypted while computations are done. Traditional encryption only protects data when stored or transmitted, but FHE enables computations without decryption. However, implementing deep learning models using FHE has been challenging due to the high computational costs. Orion addresses these challenges by offering a fully automated framework that efficiently converts deep learning models written in PyTorch into FHE programs, significantly reducing the computational load.
The framework’s effectiveness was demonstrated by achieving a 2.38x performance improvement over current methods in the ResNet-20 model, which is commonly used in FHE deep learning research. Orion also could handle much larger networks, as seen in the first high-resolution FHE object detection experiment using the YOLO-v1 model. With 139 million parameters, YOLO-v1 is much larger than ResNet-20, showing that Orion can handle real-world AI tasks efficiently.
The researchers believe that Orion’s potential goes beyond academic research and could be a game-changer for industries like healthcare, finance, and cybersecurity, where privacy is paramount. By allowing businesses to run AI models without compromising sensitive user data, Orion paves the way for more secure and private AI applications. The team has open-sourced the framework, making it accessible to developers worldwide, bringing FHE one step closer to becoming a standard tool in AI.
Further details about Orion can be read in the researchers’ preprint on arXiv.
References
- NYU Tandon School of Engineering. (2025, March 26). Encryption breakthrough lays groundwork for privacy-preserving AI models. EurekAlert!; NYU Tandon School of Engineering. https://www.eurekalert.org/news-releases/1078348
- Ebel, A., Garimella, K., & Reagen, B. (2025). Orion: A fully homomorphic encryption framework for deep learning (arXiv:2311.03470). arXiv. https://doi.org/10.48550/arXiv.2311.03470
