On September 23, Hanyang University's Department of Mathematics Professor Kim Mi-ran and her 토토사이트 위치 team announced that they have developed a new security technology that allows safe inference of large language models (LLMs) while preventing personal information leaks.

With the rapid spread of LLM-based services, the need for cryptographic technologies to protect personal information while securely operating the models has increased. However, existing homomorphic encryption–based technologies faced limitations in practical applications due to inefficient large-scale matrix operations and the need to alter the architecture and retrain the model.

To address such issues, Professor Kim's team developed a new computational method that encrypts matrices diagonally and applies ciphertext packing techniques. This allows efficient matrix operations while still in encrypted form and achieves up to a 10-fold performance improvement compared to the previous homomorphic encryption–based matrix multiplication method proposed by the team in 2018.

The team applied the developed matrix operation algorithm to LLM inference, building a framework that can run LLMs under homomorphic encryption without requiring retraining. The entire inference process is securely conducted on the computation server without decryption, and applying it to the BERT-base model demonstrated inference completion in about 10 minutes on a single GPU setup.

Professor Kim said, “This research holds significance for overcoming the speed and scalability limitations of homomorphic encryption–based matrix operations to achieve practical LLM inference,” adding, “We expect it to pave the way for safe and reliable AI services.” She also revealed, “Our recent research on homomorphic encryption–based reinforcement learning, soon to be published in Nature Machine Intelligence (IF=23.9), will further expand homomorphic encryption applications in AI.”

This study was conducted with support from the National 토토사이트 위치 Foundation of Korea(NRF), the Institute for Information & Communications Technology Planning & Evaluation(IITP), and the U.S. National Institutes of Health(NIH). The paper 'THOR: Secure Transformer Inference with Homomorphic Encryption' lists Researcher Moon Jeong-ho of Hanyang University as the first author, with researchers Yoo Dong-woo of Yonsei University and Xiaoqian Jiang of UTHealth Houston as co-authors, and Professor Kim Mi-ran as the corresponding author.

The paper's originality and excellence earned it acceptance for presentation at ‘ACM CCS 2025(ACM Conference on Computer and Communications Security 2025)’, the world's leading cybersecurity conference which will be held over five days from October 13 this year.