Fully homomorphic encryption allows computations to occur directly on encrypted data without the need for decryption, offering a robust security layer for sensitive information. Previously, the lack of uniform hardware and parameter conditions made it nearly impossible to objectively compare advancements in the field. Shruthi Gorantala, a co-lead of the benchmarking suite at Google, noted that the inclusion of Transformer inference shifts the technology from a one-off experiment to a verifiable ecosystem.

THOR functions by running the BERT language model on a single GPU while maintaining accuracy within one percentage point of its unencrypted counterpart. The framework has demonstrated significant technical milestones, including a reduction in inference time from 10 minutes to two minutes and a 9.7x acceleration in matrix multiplication. These gains address long-standing efficiency hurdles inherent to homomorphic encryption. Seungmyung Lee, CEO of DESILO, stated that these achievements are intended to be verified by the broader community rather than simply claimed, providing a transparent foundation for future development. As the company continues to participate in standardization efforts with NIST and ISO, the adoption of THOR as a baseline marks a transition for privacy-preserving AI from theoretical research to practical, industrial-grade implementation.