© 2022 Black Swan Telecom Journal | • | protecting and growing a robust communications business | • a service of |
Email a colleague |
April 2015
You admire this tower of granite, weathering so many ages. Yet a little
waving hand built this wall, and the hand that built it can topple it down much
faster. Better than the hand — and nimbler — was the invisible
thought that created the wall. Everything looks permanent until its secret
is known.
R. W. Emerson, Circles, from Essays: First Series (1841)
What an incredibly dynamic — and impermanent — marketplace we work in! Ten years ago there were no smartphones or big data platforms. Social media and the cloud were just getting started too.
So how did we adapt to all these changes? Many of you held a totally different job ten years ago. And even for those holding the same job title, the skills required of that job have probably broadened quite a bit.
The other irony is that many of the job success principles you followed ten years ago probably wouldn’t work today. And yet somehow — inexplicably — you adapted: you dove into something unfamiliar, made mistakes, and then course-corrected yourself so you could move forward.
So tell me: how can we program machines to think and act as flexibly as we humans do? That question is at forefront of R&D in fraud management systems today. And the task is devilishly hard to deliver because fraudsters are experts at devising new schemes that skirt around detection techniques — techniques that were considered state-of-the-art only a short time ago.
One person who’s thought a lot about how to move the telecom industry to a more adaptive and behavioral learning world in FMS is Shankar Palaniandy, CEO of FRS Labs, a boutique fraud solution vendor in Bangalore. For a small shop, FRS Labs is stirring its R&D ladle in quite a number of fraud management pots — IRSF detection, subscription fraud, application fraud and voice biometrics — to name a few. And in the interview that follows, he points to some of the more exciting areas FRS Labs is exploring.
Dan Baker: Shankar, your personal career required a rather flexible and adaptive path of its own. Your heritage is India, and you worked and studied in the UK for 10 years. Then there was also a one-year stint at Boeing in Seattle. But I understand these foreign business adventures didn’t really prepare you for the challenge of starting up a company in India when you returned. |
Shankar Palaniandy: It’s true, Dan. There are many different permissions and regulatory hurdles you must overcome to start a business in India. And the biggest factor of all is bribery. And all of this was a big shock to me because I had lived in the UK and US and had not seen those things.
So the trick was to navigate without having to pay a bribe. And this was especially challenging for us because the charter of our company — as a fraud management software developer — was to reduce bribery! So I could rightly say, if you ask us for a bribe, we might as well close our company.
The second big challenge was hiring people because when engineers come into a small company, they don’t see this as a viable place to work. So, during the interview if they asked me how many people we had — and at the time we had just two or three people — well, that was the end of the conversation.
That forced me to change my practice a bit. Instead of inviting them to the office, I met them at business centers and cafes to tell them about our good clients and products. But it was still very hard to convince people.
So I had to come up with a new strategy. And it was this: instead of trying to hire people from the top engineering schools, we aimed at hiring them from the mid-tier institutions instead, professionals often overlooked by big companies.
And that tactic, while risky at the time, turned out to be the right path for us. It was ideal because we had little money, and some of the engineers we hired, who are still with us, didn’t ask for a great deal of money. They were mostly asking for an encouraging work environment, good training, and the chance to do exciting and important work.
There are of course rough edges in terms of high level communication, so we spend some time reviewing emails and documents, but in terms of engineering quality, well, I think we have some of the best people anywhere — the kind of talent you would expect to find in London or Mountain View.
Learning systems is your specialty at FRS Labs. And you’ve made a mark delivering such systems to the Vodafone Group, your biggest customer. So tell me, how would you distinguish the kind of systems you build against more traditional FMS tools? |
Fraud management systems have traditionally been built as rules-based systems. If a certain threshold is exceeded, they create an alert. But our system, developed in partnership with Vodafone Group, learns by itself. It builds a profile or fingerprint of subscribers based on their actual call and handset usage: there are no rules to be built.
The system does continuous fingerprinting of a subscriber’s activity and it will pop up an alert if another subscriber comes very close to an existing fraudulent profile.
And if after reviewing that alert, the analyst decides this guy is a good customer, then that input goes back into the learning process, so the system adjusts itself to that new categorization.
But the difference is not just around writing explicit rules vs. machine-learning behaviors, our system also works on a network level, while other fraud systems work on an individual alert level. Let’s say you have three or four people working together. Our system will build up a social network profile of those people. So if a person in the group is doing something suspicious, he, himself, may not be committing fraud, but his links may actually be fraudulent.
In what applications are behaviorial learning systems most useful? |
Well, they are extraordinarily useful in subscription fraud situations.
The broken link in subscription fraud is where the fraudsters goes to a store or an online site to sign up for a service. If we can actually stop the fraud at that point, we can reduce lots of subsequent frauds such as IRSF and PRS.
For this purpose, Vodafone Ireland uses our Atreus, another fuzzy-matching and connecting mathematical modeling system. And let me briefly explain what it does.
Generally the first credit check a telecom makes when signing up a new subscriber is with a credit bureau. Identity theft is widespread today and if I steal your ID, I can easily get a phone if you have a decent credit history. Not only that, I can go to 10 different phone stores the same day, and by slightly altering the information I supply, I can obtain a phone at each store. There is enormous pressure to sell at stores, so fraud checks take a back seat — bad for business but good for fraudsters.
So what Atreus does is sit before the credit bureau check. It uses many sources of internal data — invoices not paid, bad debts written off on customers, confirmed fraudsters, suspected fraudsters, people on a watch list — and puts that all in the Atreus database. And using a fuzzy technique, it can search Dan Baker or Baker Dan or Dan Patrick or D. Baker. It also searches variations of telephone numbers, bank accounts, address, email etc. All of this in real time. So no matter what the fraudster does, it will come up with a network of matches so when the fraudster visits the second store, they will be alerted that this is the same account that obtained a phone at the first store a few hours ago.
And all of this saves Service Providers money because you reduce the number of times you run a credit bureau check. If every credit bureau check costs 5 or 10 dollars per customer and you are doing a million credit bureau dips per year, that’s a significant savings if you can shave off 20% or 30% of that cost. Also, imagine markets where reliable credit bureau data doesn’t exists — it’s a goldmine for fraudsters.
One final area I’ll ask you about is “voice biometrics”. Why do you put a lot of R&D in that area — what’s the pay-off for telecoms? |
Dan , biometrics is a very crowded space right now: lots of people are developing this kind of software. And we are focused on applying the technology in telecom fraud control.
Here’s how it works. Before we make a sale or before we activate a particular device, a SIM, or a service, we capture the spoken voice of the customer, either in agreement with all the privacy policies in place or passively when he makes a phone call to activate his service. We record the voice so we can create a voice print.
Now to do this requires very heavy mathematical models. Fortunately, plenty of research has already been done, and we use the research papers to help unpack the mathematical models to build into our product.
So if we have a voice print of Sally Smith in our files, whenever you order a new service, whether it’s a phone or broadband connection, to activate you are required to call a particular number where the voice print verifies that it is actually Sally Smith calling and Sally Smith is a good customer and not someone barred from ordering the service.
Going forward, the voice will play a bigger role in identifying you and what belongs to you. The reason voice biometrics will become more important is that mobile communications is already very extensive and the mobile technology form factor will shrink more and more to the point where wearable devices and speech input will become commonplace. So we are preparing to address the FMS needs in that new era.
Thank you, Shankar, for this fine tutorial on advanced FMS systems and learning systems. It’s been highly interesting. |
Copyright 2015 Black Swan Telecom Journal