Original title: Deep-Fake CAPTCHA: Mitigating Next-Generation Social Engineering Attacks
Guy Frankovits, Lior Yasur, Fred M. Grabovski, Yisroel Mirsky · September 2026
Scammers can use real-time deepfakes, which are AI-generated fake voices and videos, to impersonate real people during live calls. To stop this, researchers created an active defense tool called DF-CAPTCHA. Rather than just scanning media for hidden flaws, this system asks the caller to complete simple challenges. Real people can easily do these quick tasks, but current live deepfake tools struggle to fake them smoothly. The system then checks the caller on four things: realism, identity consistency, task completion, and response time. The researchers found that regular people often fail to tell real-time deepfakes apart from real callers. However, DF-CAPTCHA achieved high accuracy across both audio and video, protecting users far better than passive detection methods.
Imagine you are at a door and someone claims to be a friend. Instead of looking for visual clues, which a disguise can fool, you ask them to perform a secret handshake only your real friend would know. A real friend can do it, but an impersonator cannot. In this paper, the system does the same: it asks the caller to complete a simple task that is easy for a real person but hard for a computer to fake convincingly.
Real-time deepfake impersonation in calls is a growing threat for social engineering. People often cannot tell if a caller is real or fake. DF-CAPTCHA provides a practical defense by asking callers to complete simple tasks that deepfake systems cannot convincingly fake, protecting against fraud and identity theft.
The abstract doesn't provide specific accuracy numbers or compare against all possible deepfake methods. It also does not address pre-recorded deepfakes or situations where the caller refuses to perform the challenge. Additionally, it only targets current real-time deepfake systems, so future systems may evade it.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Deep-Fake CAPTCHA: Mitigating Next-Generation Social Engineering Attacks — Guy Frankovits, Lior Yasur, Fred M. Grabovski et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.