I know that perturbation should be proportional to the $\text{L}_1$-sensitivity of the function if someone wants ($\epsilon,0$)-differential privacy, and proportional to the $\text{L}_2$-sensitivity of the function if someone wants ($\epsilon, \delta$)-differential privacy for $\delta > 0$.

But why is that? Can anybody describe the intuition behind the use of these two norms? Or maybe share some resources?


I don't think there is a direct between the type of norm and whether $\delta=0$. For example, this crypto.SE answer mentions a mechanism that is pure $\varepsilon$-DP and scales with $L_2$-sensitivity. Conversely, some $(\varepsilon,\delta)$-DP mechanisms scale with $L_1$-sensitivity, like the truncated geometric distribution from this paper.

However, the most commonly used noise mechanisms used in DP are Laplace noise and Gaussian noise, which respectively scale with $L_1$ and $L_2$ sensitivity; the first one is pure $\varepsilon$-DP while the latter only provides $(\varepsilon,\delta)$-DP. You can find a "visual" representation of why that's the case by looking at the density function of a multi-dimensional Gaussian (top) vs. Laplace (bottom):

Density function of a 2-dimensional Gaussian distribution, with a round shape, and of a 2-dimensional Laplace distribution, with a square shape.

This picture is taken from this blog post of mine about Gaussian noise.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.