Inverting Gradient Attacks Makes Powerful Data Poisoning

Wassim (Wes) Bouaziz, El Mahdi El Mhamdi, Nicolas Usunier

This work demonstrates, in a theoretical worst-case scenario, that data poisoning attacks can mimick gradient attacks by inverting gradients. This work demonstrates an empirical upper-bound on the damage that can be achieved by data poisoning attacks. While more constrained than gradient attacks, we demonstrate that data poisoning attacks still have room for effective attacks.