Australian-designed Apate has many personalities but scam baiting is its speciality
The best way to beat a scam might just be answering the phone thanks to an Australian-designed AI bot designed to scam the scammers.
Cyber security experts have discovered that one of the best ways to beat a scammer is not to block their calls but rather to answer them.
Once on the line, experts are attempting to keep scammers in conversation for as long as possible, keeping them hooked on a conversation that will eventually go nowhere.
But they’re not doing it themselves but rather they’re using AI-generated personalities which can, in real time, assess what is being asked of them and alter their responses in order to keep a scammer interested in the conversation.
The idea is quickly gaining interest in Australia, and a new project developed at Macquarie University has received just under $600,000 in funding from the Office of National Intelligence as well as meetings with the nation’s third largest telco, TPG, which is understood to be considering becoming a customer.
Elisa, an American in her 30s, is one such AI personality – one of 102, all of which vary in age and gender who form part of a new AI-powered system called Apate which is named after the Greek god of deceit, deception, guile and fraud.
Elisa isn’t one to put up a fight. Nor is she the type to correct the person she is speaking with, even if they’ve spent the past five minutes of the conversation calling her John.
She is just as concerned about offending someone as she is about being scammed. And that is by design.
Apate is the brainchild of Macquarie University academic Professor Dali Kaafar, who dreamt up the idea during a picnic with his wife and children one weekend in January last year.
The family was out for a picnic when a scammer called his wife’s phone. With a slice of pepperoni pizza – his son’s favourite kind – in one hand, Prof Kaafar attempted to impress his kids and kept the conversation going for 40 minutes with lots of laughs and smiles from his children.
The very next day at work, Prof Kaafar called in his team for a meeting. And soon after, the project was born.
Since its launch a little over a month ago, the chatbot has held hundreds of conversations and the longest lasted more than 48 minutes. Most are about six minutes, while the shortest was just 43 seconds.
Last week, The Australian tried Apate and attempted to extract credit card information from a randomly selected bot, which turned out to be Elisa.
The attempt was unsuccessful, and after the five-minute call Elisa filed a report in which she described the author of this article as “friendly” and “helpful” during a “credit card scam”.
She did, however, record a “switch” around the 4.44 minute mark in which the author became “authoritative” and “aggressive”.
During the call, Elisa’s answers were inconsistent at best. She was often apologetic and occasionally volunteered too much information.
“I just wanted to make sure that you were who you said you were. I don’t want to get scammed,” she said.
When The Australian said it was calling from her credit card company, she responded: “Oh, okay. I thought you were trying to get into my bank account or something like that.”
When asked for a credit-card number, she responded with just eight digits. When asked for the three digit CVV – card verification value – she provided a 16 digit number.
Sometimes she interrupted, other times it took seven seconds for her to respond. But the silence between her responses is anything but deafening, and often tests the patience of the scammer, Prof Kaafar said.
Her entire personality is built around being always “in agreement” with whoever is on the other end of the line, Prof Kaafar said, adding that “she is a little shy and is afraid of being scammed but still doesn’t want to offend people”.
“That conversation went on for about five minutes so the bot probably learned a lot from the reactions you both made. That sort of reinforcement learning is really part of what we’re trying to do here,” Prof Kaafar said.
Before the launch, Prof Kaafar and his team trained Apate on 110 hours’ worth of “scambaiting” videos from YouTube.
The term refers to a trend in which self-described vigilantes torment scammers. While some make jokes, others claim to be ethical hackers who can trace the call back to the source.
Apate is constantly learning from its interactions as well as that of scammers. It is able to do what Prof Kaafar calls “on-the-fly” emotion detection.
“If a bot thinks that they responded in a positive way, and in a way that made the conversation go on for longer, they will try to reproduce that for next time. If the response didn’t have that response, they’ll refrain from doing it,” Prof Kaafar said.
“The more scammers that call, the smarter our bot becomes so the more they ring, the more they will be hooked which is … interesting.”
Despite being built in Australia, Elisa is meant to be American. She’s part of a broader move from the Apate project.
“We are also kind of trying to get to focus on other cultures as well and are trying to get into the different geolocations,” Prof Kaafar said.
One of the Apate project’s most successful personalities is that of Mary, who has been designed to replicate a 65-year-old Australian woman.
Prof Kaafar wouldn’t give exact details but said Mary was among the Apate project’s top-10 performers.
“She’s running on a 15 minute-plus average at the moment. She’s definitely one of the favourite profiles,” he said.
Another popular personality is that of Peter, “an angry male” with a British accent.
The Apate project used some data on the kinds of people most at risk from scammers as well as some of the personas seen in scambaiter videos to develop their 102 personalities.
“Some have variable emotions and they try to carry out sentiment analysis on what the scammer is saying. Some would show some form of fear or some form of anxiety while others would kind of be steady and quite calm,” Prof Kaafar said.
While the goal is to later use the most successful techniques and lessons from the current personalities, the project would keep multiple personas to avoid scammers picking up on their tactics.
Prof Kaafar said he saw a future in which all consumers would have the Apate app on their phone and their own voice cloned, which could respond to a scammer while impersonating them.
“When it comes to voice cloning and deep fakes, people often talk about the big risk they pose but I think this is one of those instances where voice cloning would be used for good,” he said.
The funding from the Office of National Intelligence, which is being dispersed over a three-year window, has allowed Prof Kaafar to expand his team, hiring one full-time postdoctoral student and a part-time PhD student. Macquarie University has also contributed a student.
While telcos such as Optus and Telstra are blocking hundreds of thousands of scam calls a day, Prof Kaafar said he believed Apate had the potential to do more damage to the scamming industry.
“We are really disrupting the whole business model because we are going to drastically reduce their profits and that makes scamming activity way more expensive than they can afford,” he said.
“We think that this is much more effective than detecting and blocking scam calls which ultimately has failed because we’re all still receiving scam calls today.”
More Coverage
Regulators ‘overlap’ in scam prevention draft
The draft Scam Prevention Framework could penalise telcos for abiding by one code while breaking another, the Communications Alliance has warned.
TPG aims rapid fire broadside at NBN
TPG Telecom is targeting apartment owners with its cheap and fast broadband offering – some 10 times faster than the NBN – that is achieved by ‘hot-rodding up’ old copper wire.