Noise explorer
Epsilon is the differential-privacy budget: a smaller epsilon adds more noise and protects more. This page bounded the noise by the column's own minimum and maximum, which is fine for exploring but would itself leak those two values. A real release needs bounds chosen independently of the data, such as public limits for the field. The re-linking figures here measure one attack on this one column; they are not the formal guarantee.
The distribution moves this much
Share of rows per value band, original against released.
Each record moves this much
Risk against information loss, every technique
Each line is one technique from weak to strong noise. Lower and further left is better. The ring marks the selected setting. The dashed line is pure chance: an attacker guessing at random. Click a point to select it, or focus the chart and use the arrow keys, then Enter.
All strengths of the selected technique
Scroll the table sideways for more measures.
Mean and spread change are relative to the original column. KS is the largest gap between the two cumulative distributions. Wasserstein / IQR is the average distance values must move to match the original, over its interquartile range. Shifts are |released − true| per record, in standard deviations.