Privacy-Accuracy Frontier

Given a dataset and a privacy level (epsilon), which algorithm(s) have the least error?

We show below (on the right) the Privacy Accuracy Frontier which is a respresentation of the best achievable error across different privacy levels (i.e., epsilon values). It plots the expected error over range queries of various algorithms on the chosen input dataset (y-axis) versus the privacy level (epsilon) on the x-axis. Each dot corresponds to the expected error of a specific algorithm at a specific epsilon. Click on a dot to visualize the noisy output of the algorithm. Dots corresponding to the same epsilon are placed along the same vertical line, and dots corresponding to the same algorithms all share the same unique color.

The points that make up the lower convex hull are those that achieve the least error at that privacy level, and are called the frontier. The points on the frontier alone can be visualized by toggling the visualize frontier checkbox on the far right.

Click the "Show/Hide Algorithms" button to view the list of algorithms. Algorithms can be included or excluded from the plot using the associated check-boxes.


The input dataset is shown to the left as a histogram of counts over a uniform grid. The noisy output of the chosen algorithm at the chosen epsilon is shown in the center. The number of bins in the output histogram matches that of the input. While the algorithms themselves may not actually generate a histogram, our visualization represents the histogram inferred from the noisy counts generated by the algorithm.

A rectangular range query can be specified on the input dataset by clicking and dragging anywhere on the input plot. The range query can be dismissed by clicking anywhere on the input. Range queries on the input are mirrored on the algorithm output. The true and noisy answers for the range query are printed below the input and output, respectively.