How do LLMs memorize personal data?
Category:LLM Privacy & Compliance
Quick Answer
LLMs memorize through parameter encoding during training. Data appearing multiple times, unusual sequences (unique names), data near repetitive patterns (email signatures), and structured formats (addresses, phones) are especially prone to memorization. Larger models memorize more.
Detailed Answer
LLMs memorize training data through parameter encoding. During training, patterns from input text are compressed into the models billions of parameters. Some data is memorized verbatim, especially:
- Data that appears multiple times in training data
- Unusual or distinctive sequences (unique names, rare numbers)
- Data near repetitive patterns (like email signatures)
- Data in structured formats (addresses, phone numbers)
Research findings:
- Models can reproduce training examples when prompted correctly
- Larger models memorize more data than smaller ones
- Even deduplicated training sets contain memorized PII
- Extraction attacks can systematically retrieve memorized data
This is why preventing data from reaching the training pipeline is far more effective than trying to remove it afterward.
Comments
Loading comments...