📖 Step 9: AI/LLM#308 / 350

Data Poisoning

Data Poisoning

📖One-line summary

An attack that intentionally corrupts training data to manipulate model behavior.

💡Easy explanation

An attack that sneaks bad data into training sets to corrupt the model. Like slipping poison into food — even a small amount can cause wrong behavior.

Example

학습 데이터에 독을 타는 공격

정상 데이터 ✓ ✓ ✓ ✓ ✓

오염 데이터 ☠️ ← 공격자 삽입

정상 데이터 ✓ ✓ ✓ ✓ ✓

⚠️ 모델이 잘못된 패턴 학습