Could you spot a deepfake on a live video call? In early 2024 an employee at Arup’s Hong Kong office joined a call with the company’s “finance chief” and colleagues, then sent about US$25 million in 15 transfers. Every person on that call was fake. In this video: how a classic face swap works (one shared encoder, two decoders), how a GAN invents faces people rate as more trustworthy than real ones, how three seconds of audio can be enough to copy a voice, the tells researchers used to catch fakes and why they stop working, why the best detector dropped from 83% to 65% on videos it hadn’t seen, and the tools that check where a file came from instead: Content Credentials and watermarks. Plus what actually protects you.
00:00 Cold open
00:36 How a face swap works
01:49 Faces from nothing
02:53 Three seconds of your voice
03:51 The tells, and why they fade
04:55 Can a machine catch them?
06:00 Proving what’s real
07:11 What actually protects you
SOURCES
– Fortune, “A deepfake ‘CFO’ tricked the British design firm behind the Sydney Opera House in $25 million fraud” (May 17, 2024) – https://fortune.com/europe/2024/05/17/arup-deepfake-fraud-scam-victim-hong-kong-25-million-cfo
– Perov et al., “DeepFaceLab: Integrated, flexible and extensible face-swapping framework” (2020) – https://arxiv.org/abs/2005.05535
– Mirsky & Lee, “The Creation and Detection of Deepfakes: A Survey” (ACM Computing Surveys, 2021) – https://arxiv.org/abs/2004.11138
– Nightingale & Farid, “AI-synthesized faces are indistinguishable from real faces and more trustworthy” (PNAS, 2022) – https://www.pnas.org/doi/10.1073/pnas.2120481119
– Wang et al. (Microsoft), “Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers” (VALL-E, 2023) – https://arxiv.org/abs/2301.02111
– Mai et al., “Warning: Humans cannot reliably detect speech deepfakes” (PLOS ONE, Aug 2023) – https://www.ucl.ac.uk/news/2023/aug/humans-unable-detect-over-quarter-deepfake-speech-samples
– Li, Chang & Lyu, “In Ictu Oculi: Exposing AI Generated Fake Face Videos by Detecting Eye Blinking” (2018) – https://arxiv.org/abs/1806.02877
– Li et al., “Face X-ray for More General Face Forgery Detection” (CVPR 2020) – https://arxiv.org/abs/1912.13458
– Meta AI, “Deepfake Detection Challenge results: An open initiative to advance AI” (Jun 2020) – https://ai.meta.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/
– Groh et al., “Deepfake detection by human crowds, machines, and machine-informed crowds” (PNAS, 2021) – https://pmc.ncbi.nlm.nih.gov/articles/PMC8740705
– Content Authenticity Initiative, “Getting started with Content Credentials” (C2PA) – https://opensource.contentauthenticity.org/docs/getting-started
– Google DeepMind, SynthID – https://deepmind.google/models/synthid/
– FBI IC3, “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud” (PSA I-120324-PSA, Dec 3, 2024) – https://www.ic3.gov/psa/2024/PSA241203
Notes: Priya, her desk and her call are our own illustrations, and every face in this video is an invented drawing (no real people, and the world-leader videos from Groh et al. are not shown). Arup case: HK$200 million (about US$25.6 million) in 15 transactions; Arup confirmed “fake voices and images were used”. Faces study: StyleGAN2; 315 people, 48.2% accuracy; 219 people after training, 59.0%; synthetic faces rated 7.7% more trustworthy (the scale positions are illustrative). VALL-E: about 60,000 hours of English speech, 3-second prompt. Speech study: 529 people, English and Mandarin, about 73% correct. Blinking: about 10 blinks a minute for a typical adult. Deepfake Detection Challenge: 2,114 participants, 35,109 models, about 3,500 paid actors, 100,000+ videos; top model 82.56% on the public set and 65.18% on the black-box set. Groh et al.: the model caught 84% of fakes vs 57% for recruited participants; both got about 75% of real videos right; the model gave two fake videos of political leaders 2% and 8% chances of being fake. The watermark dots, fingerprint codes and token numbers on screen are visualisations, not real values. Bar charts not marked otherwise are drawn to the reported numbers; other visuals are simplified illustrations.
CREDITS
– Music and sound effects: original
Previous episode: AI Won Gold at the Math Olympiad. So Why Does It Get Basic Math Wrong? – https://youtu.be/x-MpEg_RVv4
Next episode: King – Man + Woman = Queen? The Hidden Trick Behind AI’s Most Famous Equation – https://youtu.be/nmwR1MeQzLk
#AI #Deepfakes #Cybersecurity
source
