Skip to content
Advanced Academy Reader
Exit Reader

Differential Privacy in LLMs

DP adds calibrated noise to training, preventing memorization of individual data points.

advanced•2 / 6

Key Concepts

  • DP Definition: Output distributions similar whether any single record is included/excluded.
  • DP-SGD: Clip gradients, add Gaussian noise.
  • Privacy Budget: Epsilon measures leakage (lower = stronger privacy).
Section 2 of 6
Next →