A research team led by Professor Kim Mi-ran of the Department of Mathematics has developed a next-generation security technology that allows safe inference for large language models(LLMs) without the risk of personal information leakage.

Commenting on the study, Professor Kim stated, “As LLM-based services such as ChatGPT are rapidly expanding, there is a growing need for encryption technologies that can both protect personal data and ensure reliable system operation. This study aimed to address security challenges while significantly improving performance.”

Existing homomorphic encryption(HE)-based techniques have long faced practical barriers due to inefficiencies in large-scale matrix computations and the need for model restructuring and retraining. To overcome the barriers, the team proposed a new computational approach combining 'diagonal matrix encryption' and 'ciphertext packing,' which allows multiple 토토사이트 매입 sets to be stored in a single cryptogram.

Professor Kim explained, “We stored the matrix diagonally within ciphertexts and combined addition, multiplication, and rotation operations to implement efficient matrix computation. This approach enables rapid processing even while data remains encrypted.” The new technique demonstrated over ten times faster performance compared to the team’s previous matrix multiplication framework developed in 2018.

Applying this algorithm to LLM inference, the team built a robust framework that executes models securely within a homomorphic 토토사이트 매입 environment without the need for retraining. In BERT-base model(with approximately 110 million parameters) experiments, inference using a single GPU in about ten minutes was proven possible.

Professor Kim noted, “While the conventional NEXUS method required roughly 2.7 hours for BERT-base inference, our technology reduced this to under ten minutes—a more than sixteenfold improvement.” She added, “We optimized operations in the inefficient multi-head attention layer which required parallel matrix calculations from previous studies, enabling seamless iterative computations without redundant data transformations.”

The paper was recognized for its originality and technical excellence and was presented at ACM Conference on Computer and Communications Security(ACM CCS 2025), one of the most prestigious global conferences in cybersecurity. The study also received the Grand Prize at the 2025 National Cryptography Contest, hosted by the Korea Cryptography Forum and sponsored by the National Intelligence Service(NIS).

A broad adoption of the developed technology in fields requiring strict personal information protection, including healthcare, finance, and public data management is anticipated. “When using cloud-based LLM services, users must provide personal data,” said Professor Kim. “By implementing this technology, inference can be safely processed without decrypting sensitive information.”

Addressing recent large-scale data exposure incidents, she emphasized, “The root cause lies in companies’ lax data management practices. Going beyond basic encryption, organizations must adopt technologies that enable computations while encrypted.” With this approach, services can maintain full security without the need for decryption, ensuring protection of personal data at its core. “This study lays the technological foundation to fundamentally resolve the issue of personal data leakage,” she added.

The team is now extending the technology to enable efficient inference in larger models such as BERT-Large(with approximately 340 million parameters) and Meta’s LLaMA(Large Language Model Meta AI). “In the long term, we are exploring ways to apply homomorphic encryption-based computations even during the training stage,” said Professor Kim.

Highlighting its practical significance, she mentioned, “Functions like ‘Generative Photo Edit’ in Samsung Galaxy models operate in a cloud environment, requiring users to upload images through networks. Since such processes involve external servers, reinforced privacy-preserving AI technologies are essential.” Her team plans to further advance LLM-based technologies that ensure privacy protection in cloud-driven, personalized services.