#South Korea #Seoul #Thor #Desilo #FHE

DESILO's Groundbreaking THOR Achieves Global Benchmark Recognition

In a remarkable advancement for encrypted artificial intelligence, DESILO Inc. has announced that its innovative framework, THOR, has been designated as the reference implementation in the global Fully Homomorphic Encryption (FHE) Benchmarking Suite. This framework is groundbreaking as it is the first to allow for the entire process of large language model (LLM) inference to be conducted directly on encrypted data without needing to decrypt it, thus ensuring maximum data privacy and security.

The Significance of THOR in the FHE Benchmarking Suite

The inclusion of THOR in the FHE Benchmarking Suite represents a significant milestone in establishing a universal benchmark for private AI and encrypted LLM inference. This suite enables organizations across sectors such as healthcare, finance, and government to deploy powerful AI models safely, without jeopardizing sensitive information.

Fully Homomorphic Encryption (FHE) has emerged as a crucial technology, facilitating computations on encrypted data while preserving privacy. Yet, the landscape has faced challenges where previous results were often reported under varying conditions regarding hardware and parameters. This inconsistency made it difficult for stakeholders to assess progress objectively. The FHE Benchmarking Suite seeks to address this issue, creating standardized, reproducible workloads and performance metrics to promote transparency in the field.

According to Shruthi Gorantala from Google, who co-leads the benchmarking initiative, “With this structured approach, encrypted inference transitions from being a series of isolated demonstrations to a reproducible and comparable result, essentially laying the foundation for an ecosystem.”

Performance Highlights of THOR

At the heart of THOR's success is its ability to run the esteemed AI language model, BERT, while keeping the data encrypted throughout the process. This framework allows for an end-to-end computation flow—from input through to output—without any need for separate retraining, thus maintaining high levels of accuracy similar to an unencrypted counterpart (within one percentage point).

Significant performance improvements have also been observed:

• - Drastic Speed Improvements: Inference time has significantly reduced from approximately 10 minutes to just 2 minutes due to ongoing optimizations.
• - Matrix Acceleration: THOR enhances matrix multiplication performance—critical to the computation—by as much as 9.7 times compared to the previous benchmarks, effectively tackling one of the most significant challenges associated with FHE.
• - Universal Standardization: The methods and results from THOR can be validated and reproduced under standard conditions, offering a baseline for evaluating future encrypted language-model inference technologies.

The Vision for Private AI

Seungmyung Lee, CEO of DESILO, emphasized the importance of collaborative progress in this field: “By establishing a common foundation for encrypted language-model inference, we can ensure achievements are verified rather than merely claimed.” This pioneering step not only illustrates DESILO's commitment to advancing Fully Homomorphic Encryption but also their active participation in discussions surrounding standardization at both the U.S. National Institute of Standards and Technology (NIST) and the International Organization for Standardization (ISO).

As DESILO continues to innovate and push the boundaries of encrypted AI for practical enterprise and public sector applications, THOR stands as a testament to the potential of secure data processing and privacy preservation, heralding a new era for the global tech ecosystem. The implications for industries managing sensitive data are monumental, illustrating a future where AI can be leveraged to its fullest potential without compromising security or privacy.