We’ve seen a rise in "vishing" attacks where executives' voices are cloned. Can we utilize Generative AI (ChatGPT, Gemini) to build better detection agents, or are we entering an era where we can no longer trust voice or video for identity verification during high-value wire transfers? What are the new protocols for remote identity proofing?
3 answers
The threat of deepfakes is terrifying for the financial sector. Using Generative AI (ChatGPT, Gemini) to analyze audio for "unnatural" artifacts is one approach, but it's a cat-and-mouse game. We’ve moved back to a "code word" system for any transaction over $50,000. It sounds old-fashioned, but an out-of-band physical verification or a pre-shared secret that isn't stored digitally is currently the only way to be 100% sure. We are also looking into blockchain-based digital signatures for video calls to ensure the stream hasn't been tampered with in real-time by a generative model.
Is there any specific software you recommend for real-time detection of synthetic audio during standard corporate Zoom calls?
We updated our policy to require a "liveness test"—asking the person to turn their head or hold up a specific object—during unexpected video calls.
Liveness tests are great! Most current real-time deepfake models struggle with profile views or rapid background changes, so that’s a very smart, low-tech fix.
Douglas, there are a few emerging tools like Pindrop that specialize in this. However, they aren't perfect. We find that simply training staff to look for "lag" or inconsistent lighting in video calls is still a very effective first line of defense.