Skip to main content
Skip to content
Advanced
Academy Reader
Exit Reader
Reset
Differential Privacy in LLMs
DP adds calibrated noise to training, preventing memorization of individual data points.
advanced
•
5 / 6
Evaluation
Metrics: Epsilon via accountants; utility via downstream tasks.
Trade-offs: Noise degrades performance; scale data to mitigate.
← Previous
Section 5 of 6
•
Tip: Use ← / → to navigate
Next →
Back to top