Part V · Chapter 53Version 1.1 · 5 November 202635 min read
Adversarial Debiasing
Key message
Adversarial debiasing trains a predictor to be accurate while a second network fails to recover the protected attribute from its outputs or representations; it can enforce demographic parity or equalized odds, but only an independent leakage test shows it did.
Five chapters or files are free. This one would be your sixth.
Leave an email address and every chapter and every PDF of the compendium opens, in this browser, for a year. Nothing else is asked, and whether we may contact you is your choice.