Anonymisation tries to keep data useful for analysis while making it impossible to tell who is who.
It matters because truly anonymous data falls outside most privacy laws, while identifiable data does not.
Doing it properly is much harder than it looks.
Direct identifiers, such as names, emails and phone numbers, are removed or replaced first.
Quasi-identifiers, such as postcode, age and gender, can still identify people in combination.
Generalisation coarsens them, such as turning ages into 10 year bands.
Suppression removes values or records that remain unique after generalising.
K-anonymity is reached when every combination of quasi-identifiers is shared by at least k records.
It can still leak if everyone in a group shares the same sensitive value, such as one diagnosis.
Linkage attacks match a released dataset against public records, such as electoral rolls, to find people.
Pseudonymisation replaces identifiers with codes, which helps but can be reversed by anyone holding the key.
Differential privacy adds random noise to results, with a budget that limits how much any person can affect them.
More protection means more noise or coarser data, so there is always a trade-off with usefulness.
In practice, sharing totals and averages instead of rows is often the simplest way to reduce risk.
The usual mistake is treating removal of names as anonymisation, when combinations of ordinary columns can still identify people.