Executive Summary
In a recent study, Microsoft found that 25 out of 28 firms surveyed did not have protections against the so-called adversarial attacks on their machine learning (ML)-based systems.* One of the banks surveyed responded:
“we want to protect client info, employee info used in ML models, but we don’t have a plan in place.”
Adversarial ML is concerned with malicious attacks against ML models. The main goal of adversaries is to trick machine learning models by providing specialized, deceptive inputs that purposely confuse an ML model.