Easy-to-interpret, unified, tunable bounds on major operational attack risks in privacy-preserving ML and statistical releases that are more accurate than prior methods, using f-DP