mncai/Korean-PII-Masking-Model

Korean PII Masking

Author: 임성준 (Sungjun Im) [LinkedIn]
Role: Project Lead & Primary Researcher

Overview

BERT-based token classification model specialized for Korean Personally Identifiable Information (PII) masking. It detects and masks 14 common types of Korean PII entities.

For best accuracy in production, strongly recommended: use a hybrid approach
(pre-processing → model inference → post-processing (Regex and rule based) rather than the model alone.

Base Model & Architecture

• Base pretrained model: KcBERT-Large
• Model type: BertForTokenClassification
• Architecture highlights:
  • Hidden size : 1024
  • Layers : 24
  • Attention heads : 16
  • Intermediate size : 4096
  • Max position embeddings : 300
  • Vocab size : 30,000
  • Activation : GELU
  • Dropout : 0.1 (hidden & attention)

Supported PII Types (BIO tagging)

1. 가맹점명 (Business Name)
2. 결제금액 (Payment Amount)
3. 계좌번호 (Account Number)
4. 로그인ID (Login ID)
5. 상세주소 (Detailed Address)
6. 신용점수 (Credit Score)
7. 여권번호 (Passport Number)
8. 우편번호 (Postal Code)
9. 운전면허번호 (Driver's License Number)
10. 이름 (Name)
11. 전자메일 (Email)
12. 전화번호 (Phone Number)
13. 주민등록번호 (Resident Registration Number)
14. 카드번호 (Card Number)
15. 휴대전화번호 (Mobile Phone Number)

Example

입력: "양철용 고객님, 8월 10일 14:32에 백다방 코엑스점에서 9,910원 결제 내역 확인됩니다."

출력:
- 발견된 PII:
  - 양철용 -> [이름]
  - 백다방 코엑스점 -> [가맹점명]
  - 9,910원 -> [결제금액]

This list focuses on the most frequently occurring and sensitive personal data types in Korean text/documents.