Digital Doppelgängers: The Dual Faces of Deepfake Technology

The session opens with a hypothetical scenario in which a security analyst receives an urgent request, seemingly from a manager on a video call, to loosen network controls. From there the speaker explains how generative adversarial networks pit a generator against a discriminator, then sorts synthetic media into text, images, audio and video. Each category gets a concrete example: fabricated legal citations that a court rejected, a romance scam built on generated profile photos, and a robocall imitating a national leader.

The middle of the talk moves into live demonstrations, including a voice clone built from a short audio sample and a discussion of real-time face-swap tools running on webcams. Case studies cover a Hong Kong finance employee who wired about $25 million after a video call with what turned out to be fake colleagues, and a company that nearly hired a North Korean operative using a face-swapped identity, whose laptop was shipped to a laptop farm.

On detection, the speaker finds audio classifiers improving faster than video tools, which produced false positives and misses in testing. The practical defenses centre on people: asking unscripted questions, requesting a quick head turn, and confirming sensitive requests through a separate, known channel. Organisational steps include an AI-content proxy, clear policy, and frameworks such as the NIST AI RMF and MITRE ATLAS. The takeaway is to verify before acting and to keep a healthy level of skepticism.