The Privacy Loss Parameter (ε)
What you'll learn: How the parameter epsilon (ε) quantifies privacy guarantees in differential privacy mechanisms.
Understanding Epsilon
In differential privacy, ε (epsilon) is the single number that tells you how much privacy you're sacrificing. Think of it as a "privacy budget" — the lower the epsilon, the stronger your privacy protection.
Epsilon mathematically bounds how much an adversary can learn about any individual from your model or dataset. Specifically, it limits the maximum ratio of probabilities that any output could be produced by two datasets that differ by exactly one person's data.
The Privacy Spectrum
- ε = 0: Perfect privacy (but useless — you learn nothing from the data)
- ε ≈ 0.1 to 1: Strong privacy guarantees, suitable for highly sensitive data
- ε ≈ 1 to 10: Moderate privacy, commonly used in practice
- ε > 10: Weak privacy, approaching no meaningful protection
Intuitive Analogy
Imagine epsilon as the "blur level" on a photograph of sensitive information. A very small ε is like heavily pixelating the image — you preserve privacy but lose detail. A large ε is like a light blur filter — you can see most details, but privacy protection is minimal.
Why Epsilon Matters
When implementing differential privacy in ML, you must choose ε before training. This choice reflects your trade-off: smaller epsilon means stronger privacy but typically lower model accuracy or utility. Organizations must balance regulatory requirements, ethical obligations, and model performance when selecting this parameter.
Key Takeaway: The privacy loss parameter ε is the fundamental measure of privacy in differential privacy — smaller values provide stronger mathematical guarantees that individual data remains protected, at the cost of model utility.