One of the classic applications of machine learning is in the area of fraud detection. We often read about the utility of ML when it comes to, for example, spotting credit card fraud or other anomalous behavior in the digital realm. This week, however, security company Brivo announced they would be applying this technology to the work they do to protect the physical world, as well. In their announcement, Brivo notes the big changes to office-use patterns that have occurred since the pandemic, and the challenges this poses to human or traditional security measures to actively guard against unwanted access or intrusion, both to those physical spaces and the valuable digital property that make up a company. Their application of machine learning is an interesting new twist on a relatively standard use of this tech.
In general, systems to detect fraud that are not powered by ML suffer from a significant number of challenges. It can be very cumbersome to try and imagine all of the possible events and interactions that may indicate something untoward is taking place, but each of those outcomes must be carefully mapped out for a computer to identify them and intervene. For example, perhaps if someone logs in to his/her account between 2:00am and 5:00am that is considered normal, but if they log in during those hours and they attempt to execute a transaction that is 250% or larger than the average of their last 10 transactions, then it is likely fraudulent. As you can imagine, developing a list of highly particular possibilities would be incredibly difficult and likely to miss much fraudulent activity, while also over classifying normal behavior as fraudulent.
With machine learning, however, past data can be considered on a massive scale and identify potentially fraudulent activity from its own learning. Rather than programming every possible outcome, the machine’s model will allow it evaluate every event or pattern of events and determine if, based upon past behaviors, it appears fraudulent or not. It does this through a process of supervised learning, during which a human feeds lots of data into a computer and shows it examples where the outcome of the historic behavior was “normal” versus “fraudulent” and the computer then learns how to spot the difference between the two in the future. Unsupervised learning can also be deployed in this context. With unsupervised learning, the machines are trained simply to spot anomalous behavior, which can be helpful in flagging potentially new types of fraudulent activity that the supervised models may not yet be trained to recognize. Once humans confirm the anomalous behavior is, in fact, a new type of fraud, that info can be used to update the supervised model so it can learn to spot similar behavior going forward.
Back to Brivo. The ability to leverage this technology in the physical world is a great example of how even these classical ML tools are finding new life and fueling a growing wave of use cases. As we’ve seen elsewhere, the proliferation of digitized, structured data creates new opportunities for the application of machine learning. It wasn’t long ago that office doors were opened with physical keys, but even more modern access tools (i.e., RFID fobs and ID cards, for example) were not necessarily created with the intent or ability to generate a comprehensive database of user activity. Companies like Brivo have aimed to change that, by creating a unified picture of a person’s online and offline behavior, when it comes to everything from entering the parking garage to logging into a secure user portal. By doing so, they have created the opportunity to use this data to fuel machine learning models that are able to consider both physical and digital world datapoints when evaluating potentially problematic behavior.
The Brivo pattern is one that seems increasingly common for tech-enabled companies. They have brought digitization and automation technology (among much else) to industries. By doing this they have caused massive disruption and introduced significant convenience and efficiency. But a new type of innovation increasingly tends to follow-- either from a fast-moving incumbent, or from a scrappy challenger. That new innovation builds upon the newly digitized space to build clean, structured data that can then be used to power intelligent machines. This is a pattern that is shaping up to be very influential, and the ability of businesses to take advantage of it will likely bear significantly on their success over the next decade.