How we could prevent poisoning attacks in adversarial Machine Learning?

I read it from this link and other sources. As per my understanding, poisoning could be done after the ML algorithm has been made or while building up the model with test data. One example I read was like a car is driving and a small image could be pasted on a wall, which could make it turn left, so the car's AI algorithm misclassifies it.

But for poisoning the test data the attacker needs access to internal software at the time before the model is built so that the model that is built is corrupted. How could an attacker do that? That seems impossible. Or it could be in cases where the ML model is being built dynamically.

Just poured my thoughts out. I am interested in knowing thoughts about the above, and, specifically, what are the ways in which poisoning could be prevented?



You must log in to answer this question.

Browse other questions tagged .