How to Verify Real People in AI Calls
We are so used to seeing and hearing people from all over the world in video conferencing and telephone calls that it has become a norm to identify individuals by their voice or face.
This has traditionally been the way we verify people in phone calls and video meetings. Recently, however, technology has enabled the creation of realistic fake faces and voices for all sorts of nefarious activities. So how can we verify people in these situations? The answer is simple: a shared secret chosen beforehand.
When seeing and hearing is no longer proof
In addition to stealing identities and committing financial identity theft, deepfake technology can also be used by criminals to conduct voice-over-the-phone scams and video conferencing scams. For example, all that a scammer would need to carry out a convincing voice-over-the-phone scam is a short audio clip of a person’s voice that was recorded in some way (e.g., posted online, recorded at a webinar, in an interview, meeting, etc.). Similarly, for a video conferencing scam (e.g. mimicking a meeting with real people), the scammer would only need a deepfake video, created from previous video footage of the real people (posted online, presented at webinars, in interviews, meetings, etc.).
All this software can do is replicate information that has been previously recorded or posted (e.g., pictures, movies, audio files). Therefore, by definition, such software will never be able to generate information with respect to secrets that have never been written down or recorded in any shape or form (i.e. never been sent by email or text message, never posted on the internet, etc).
A strong security question, or a family safe word, or a workplace verification question, all are very important today to verify people who you don’t know personally.
At Home: Voice Clones
Here is an example of what that call would look like.
Scammers will use AI to create audio that sounds like your children or grandchildren, as they will often use panic to try and get money released quickly and asap for their troubles. The panic that is heard in the audio will sound very real and can be heartbreaking.
A scammer can use AI to make voice calls that sound like your loved ones. Then they claim they are in trouble and you must send them money immediately. These scams are known as family emergency scams. The FTC warns that scammers can use AI to clone a loved one’s voice for these scams. The FBI advises families to create a secret word or phrase to verify identity. A family safe word is easy to set up. The word should be something that only your family knows. Then on the scam call, ask for the word. If the caller cannot give you the word, hang up. Then call the real person back on a number you already have.
A voice clone would have no idea of your private words and passwords which have not been recorded, posted, emailed, or texted.
If the would-be scammer cannot provide your family’s safe word, hang up and contact the family member or parent using a previously established contact method or time.
At work: the $25M deepfake video call
This single deepfake video conference led the employee through what seemed to be a normal process for transferring a confidential amount of funds to various bank accounts. He was reassured by what seemed to be the CFO and other senior colleagues on the call. In total, the employee sent 15 transfers worth around $25 million before realizing what had actually happened. The incident was covered by CFO Dive and various other media outlets.
They were AI-generated fakes.
Reassured that senior colleagues from around the business had been making the arrangements, the employee transferred a total of about $25 million across 15 payments. A detailed account of the scam and the aftermath has been reported by CFO Dive and others.
Recognizing the face of someone calling is not enough to verify the person. Verifying a person to verify transactions of money, access or data that are sensitive in nature takes more than recognizing the face of someone calling.
Using shared secrets to verify people
Verifying someone’s identity during a phone call works much the same way as creating a security answer to use during very stressful times. Setting up a family’s safe word for verification during calm times will help to prevent scams.
-
Family safe word: Arup’s experience with a $25M deepfake incident highlights the importance of a family safe word. Choose a word that has never been published online and is unusual enough that it would not normally be discussed in public. Agree on it with family members in person. On any call where you’re being pressured to do something, ask for the safe word, then call the person back on a phone number you know they use.
-
Create a verification question for money and access. Think of a secret question and answer for wire transfer approvals, for changing payment details, for accessing sensitive information (even if it’s within the company). The question and answer must be something that no one will be able to find out. Verify this information via a different communication channel (e.g. a phone call) before taking any action.
-
Check the person, not the screen. If someone calls you claiming to be from IT or another department, check their name against the company directory to see if it matches up with the person who called you. Then call them back from a number you know is theirs. (Note that an actual emergency will not mind waiting a few minutes for you to verify their identity.)
-
Treat a safe word like any other secret answer. This secret should remain offline. Only share the safe word with people who need to verify you. Never distribute a safe word and then forward the distribution email to others later.
Questions people ask before they set this up
Short answers for account recovery, family safe words, and identity checks at work.
Why do security questions exist in the first place if they are so weak?
Many online services use security questions as a secondary form of authentication, so it is better to set up good security questions and answers for your online accounts rather than to not have any. Setting up weak security questions and answers for your online accounts can actually decrease security rather than increase it.
Even if security questions are poor for account recovery, that does not mean that you should answer them poorly for your online services. The weak spot in security questions is the truthful answer. Your real birth city is probably already easily found online.
Scammers can generally look up a truthful answer to a security question. A made-up answer that only you know is much harder to use against you. Treat the answer to your security question like another password. Create it, save it somewhere secure, and enter it exactly as you saved it.
Do I have to answer them truthfully?
Usually, no.
Remember, the service does not care if the answer you provided for a question like “city where you were born” is a real city or your actual birth city. The service is trying to verify that you can recall an answer that you provided at a prior time. Treat answers to security questions like you would additional passwords. Create an answer, and store it securely so that you can recall it and enter it exactly each time.
A great password is something that you only know. Writing it down and keeping it in a secure place, such as a home safe or password manager online, is actually better than trying to remember it.
What makes a good safe word?
A two-word phrase where the two words are not naturally combined together to form a word, such as “copper kayak”.
Words that are normally found together should be avoided. For example, a person’s Facebook or LinkedIn page, etc, may contain a phrase of two normal words that you have chosen as your safe word. Two normal words that do not normally go together (i.e. a phrase) would be better. Something like ‘copper kayak’ could be perfect. Then it will be easy to remember and say out loud, and you can share it with family and/or carers as required.
Someone on a video call is urgently asking me to move money.
Now what?
Pause the request.
Always have a second way of verifying someone. That could be an agreed-upon verification question, or an alternative method of verification such as calling another known number, checking the internal company directory, following an approved workflow, etc.
Even if you do confirm in a callback call, it won’t have cost you anything to check in the first place, and it’s the pressure to skip the check in the first place that is the warning sign.
Can I spot a deepfake by looking for glitches?
Do not count on it.
Deepfakes are continually getting better, so it’s unlikely you’ll be able to verify a video as fake in time to deny a payment or account access. Instead, verify the real person behind the video using a shared secret or a callback that you trust.
How often should I update these?
These details should be updated as needed for any security breach reported by the service, if you suspect an answer may have been exposed or become easier to discover. A family safe word should only be updated if someone outside of your trusted group has heard the word.
A family safe word is usually enough and does not need to be changed. But if people outside of your trusted group might know the answer then you should change it.