Privacy protection where removing one person's entire dataset (all their samples) produces negligibly different algorithm outputs.