- A song streams from your Pandora app, and you click “thumbs up!”
- A parent reads a book to a baby, taps an image of a Labrador and says, “dog!”
- A student looks at the results of a Google search and clicks the fourth link.
What do all three of these things have in common with a data scientist's work? And why should an InfoSec professional care?
Software applications and babies can’t read your mind—but with your help, they can recognize
patterns, and adjust their behaviors accordingly. To Pandora, music is data, and a particular song is made up of many different features, including tone, beats per minute, instrumentation, and many others (see the Music Genome Project). When you click the Pandora "thumbs up," you help give that complex constellation of features a simple label—in this case you are saying, “I like it!”
This is called “labeling,” and it’s a vital part of training Artificial Intelligence (AI) across many domains—including in Cyber Security.
Labeling allows AI to understand the implications of a pattern. Once an AI knows the label associated with a pattern, it will evaluate future patterns based on that knowledge. When you “thumbs up” a song, you are helping the algorithm find other songs that share some of the same features. If you continue to click “thumbs up” on the next set of songs, Pandora becomes even more precise in finding similar songs. This also applies to AI in a Cyber Security scenario: labels “train” AIs to recognize attacks, while bypassing innocuous interactions.
This training occurs both ways. Let’s think about what happens when an AI is incorrect—Pandora sends a listener a song they don’t like, a baby misidentifies a cat as a dog, or Google comes up with three top results that are irrelevant to the original search.
What action does the human take? In Pandora-land, the human clicks thumbs-down. Or the parent corrects the child. Or the student clicks the fourth link displayed by the search. Now a new label has been generated, and that helps the AI (or baby) become more precise in identifying what the human will call the pattern the next time it observes it.
The same goes for InfoSec. If an AI classifies an innocent interaction as harmful, or vice versa, an expert can correct it. As a result, it gets a little more intelligent—it will never make that same mistake again. And unlike you when you’re trying to figure out something new, AIs don’t need sleep or lunch breaks. They can learn 24/7/365.
So next time you’re grooving along to a song in Pandora, spare a thought: you’re not just getting cool tunes from this friendly AI—you’re making it smarter, too.