How can businesses detect real-time video deepfakes during high-stakes corporate video calls?

All answers to this question.

Detecting "live" deepfakes requires a mix of technical tools and human observation. Look for "glitches" or artifacts around the edges of the face and neck, especially when the person turns their head. Often, the AI struggles with rendering consistent lighting or realistic eye blinking. Technically, you should look into liveness detection software that analyzes skin textures and micro-expressions that GANs often fail to replicate perfectly. We’ve started requiring "live challenges" where the speaker must perform a specific action, like holding up a hand or turning sideways, which often causes real-time deepfake models to lose alignment or blur significantly.

   Answered 2024-03-22 by Linda Thompson

Beyond the visual checks, have you considered implementing a secondary "out-of-band" authentication method, like a hardware security key or a pre-shared verbal code, before authorizing any financial transfers?

   Answered 2024-03-28 by James Henderson

• James, that is a great suggestion. We actually implemented a "duress word" policy last month. If a request feels off, the employee asks a seemingly random question that requires a pre-shared answer. Since the deepfake is usually controlled by a script or a basic LLM, it can't always pivot to personal, non-public knowledge quickly. It’s a low-tech solution that perfectly complements the high-tech detection software Linda mentioned.

     Commented 2024-04-05 by Mark Sullivan

Look for inconsistencies in the background. Often, the deepfake is crisp but the background has strange warping or shadows that don't match the speaker’s movements.

   Answered 2024-04-15 by Sarah Miller

• Sarah is right. These "boundary artifacts" are the most common giveaway in current real-time generative models because they struggle with edge-blending.

     Commented 2024-04-20 by Linda Thompson