Protecting AI

- [Instructor] Now that we've talked…about the different ways that AI can be attacked…I want to say a few things about how it can be protected.…As a quick review there are two kinds of attacks.…The evasion attack which is an attack…on an implemented model,…and to attack this you have to create new examples,…new images, or new audio recordings…that fool the algorithm.…Now the nice thing…is those can be potentially defeated…by updating and improving the algorithm…so it doesn't have these very specific susceptibilities…that were being exploited.…A poisoning attack on the other hand…is an attack on the model in development…and that has the unfortunate consequence…of baking in the mistakes into the algorithm…and frequently updated algorithms are especially vulnerable.…

That's problematic because…one of the great virtues of artificial intelligence…is that it's frequently updated,…that it can develop as it gets new information…and learns from it's old errors…and so this makes things especially complicated.…But before we look at what can be done about it,…