From the course: Google Cloud Professional Machine Learning Engineer Cert Prep

Data poisoning - Google Cloud Platform Tutorial

From the course: Google Cloud Professional Machine Learning Engineer Cert Prep

Start my 1-month free trial Buy for my team

Data poisoning

“

- [Instructor] Let's take a look at some insider threats that could occur via data poisoning. This is often not looked at in terms of a threat as what could an insider do into your organization and this is why it's important to carefully consider the concept of principle of least privilege. So let's say in this first attack vector here, we have a corporate-owned system, let's say Google Drive, email, et cetera. And we know that there's a particular type of image that could cause your Google Drive to go offline. So what could happen is that someone inside a corporation could willfully put a prohibited image inside of the Google Drive which would then make sure that the corporation is offline. And it would take some time for that organization to be able to contact Google and explain exactly what happened. So this could be an extended "denial of service" attack on a corporation, as the corporation tries to turn the Google assets back online. So this is actually pretty sneaky, but easy to do, depending on what privileges a employee has in an organization. They know that, in fact, they could even put something that is very hard to find out that it's a prohibited image. Maybe it's a World War II image or something like that. Now, a second one that is an attack vector is that potentially an insider could secretly seed training data with images that are known to trigger some kind of forbidden attack. And so in this case, they could plant a file on a customer system and that system would train the data and eventually the multi-class classification model would see something that wasn't really there. They could plant essentially a false flag of forbidden content. And that particular organization could constantly get a denial of service attack because an insider was explicitly trying to take out another organization. So in both cases here, the training data that's planted, or in the case of the prohibited image, one of the ways that this could be mitigated is through regular auditing, reducing the principle of least privilege approach. So making sure that people don't have more access than they need to. These are not theoretical attacks. It's very important to consider data poisoning as a real threat to many organizations as data and email gets more popular.

Contents