Exchanging Scale and Epsilon

The scale of the data and the epsilon privacy parameter contribute to error in equivalent ways.

Because we always report scaled error, error rates of all algorithms decrease with increasing scale. Error rates also decrease with increasing epsilon. Intuitively, greater scale (i.e. more data) and greater epsilon (i.e. more privacy budget) both increase the "signal" available to the analyst and result in lower error. For virtually all algorithms we consider, these two factors have precisely equivalent effects on error. Scaled error is determined by the product of scale and epsilon, and for experimental evaluation we do not need to vary scale and epsilon independently.

Scale-epsilon exchangeability

As a concrete example, we demonstrate scale-epsilon exchangeability in the context of the Laplace mechanism and the error of a single counting query.

Error varying with scale

Notice that at each scale, increasing epsilon has a predictable reduction in error. Across scales, we see equal error when a change in scale compensates for a change in epsilon. For example, the following result in the same error:

  • (epsilon=1.0, scale=1,000)
  • (epsilon=0.1, scale=10,000)
  • (epsilon=0.01, scale=100,000)

Error varying with scale-epsilon product

The above relationships are shown more clearly when we plot the same set of relationships by scale-epsilon product (on the x-axis).

  1. Remove "L2" from y-axis label(s) (don't think we need it in this case)
  2. In the second plot, the scale is implied but technically ambiguous. Maybe a mouseover should show the scale? (not that critical) (What do you think? I could make them mouseover.)