Exploring Efficient Implementation of Language Models Under Homomorphic Encryption

LIST

모드선택 :              
세미나 신청은 모드에서 세미나실 사용여부를 먼저 확인하세요

Exploring Efficient Implementation of Language Models Under Homomorphic Encryption

김한나 0 282
구분 박사학위 논문 발표
일정 2025-06-19(목) 13:00~14:00
세미나실 27동 220호
강연자 노동환 (서울대학교)
담당교수 강명주
기타
This thesis focuses on the efficient implementation of language models under homomorphic encryption (HE). Recently, as people increasingly use large language models (LLMs) and receive personalized answers, privacy concerns have arisen. Accordingly, privacy-preserving machine learning (PPML) has attracted growing attention alongside the development of LLMs.

However, implementing LLMs remains prohibitively slow due to the notoriously high computational overhead of HE. This is mainly because theoretically HE only supports addition, multiplication, and rotation.

In this thesis, we present methods for the efficient implementation of LLMs under HE, particularly in the CKKS (Cheon-Kim-Kim-Song). First, we modify the attention mechanism and explore the previously unexamined effect of LoRA (Low-rank Adaptation) for speedups of fine-tuning and inference of LLMs under HE. Second, we propose a novel and efficient next-token prediction algorithm which has not yet been developed under HE.

프린트
    정원 :
    부속시설 :
세미나명