The next target for hackers
I was speaking recently with Eric Siegel, PhD about machine learning and predictive analytics, specifically what he likes to call The Power to Predict Who Will Click, Buy, Lie, or Die. In that conversation, we talked about the black box that lies at the heart of many company’s AI deployments. But we didn’t talk about the security considerations that stem from this situation: that is, it’s not just a company’s computers that are vulnerable to hackers. It’s the AI running on those computers, that’s making use of data.
Every day, we hear about companies deploying applications with the dreadful acronym AI to provide the best customer service or something. They talk up their solution, the most advanced solution, with the best AI that money can buy. And of course all that AI that is their big differentiator is a highly guarded secret. It’s also a source of vulnerabilities.
And this is where it gets tricky. In order to build your analytics, you need data. You need to take the data and build models. Where can you get the data? If you are a mature organization, the chances are that you already sit on a lot of it. In the case of a startup, most likely you go on the Internet and find publicly available data sources like this one, download it and start training and building your models.
As you can imagine, that data set can be anything from finance, medicine, logistics, but also language (either text or voice) or images; and it can contain millions of objects. It is already difficult to obtain the data, clean it and get it ready for processing. Trying to identify which information inside is real or fake, is much more difficult - or due to resource limits, actually impossible.
To make things even worse, organizations are declaring any algorithm a trade secret. They provide no visibility into the inner workings. Sometimes, this is done because the makers of these algorithms can't even explain how it works!
Maybe you heard the term AI Explainability. It is an emerging field to define transparently how the machine arrived at any particular decision. Unfortunately, few companies do this, even if they should.
What can you do? Demand, when anytime anyone mentions AI or ML that they explain to you, in a language you can understand, how it works and how it was trained. In the case of an image recognition system, ask them if their algorithm can tell the difference between a Chihuahua and a muffin.
The recurrent pattern here? In order to take advantage of all the new wonders we are building, we have to keep learning to stay in control, otherwise we end up like sheep who have no say about the future.