3,839
edits
Synthetic human-like fakes: Difference between revisions
Jump to navigation
Jump to search
→Digital sound-alikes: mv contents + change subheading to === Documented crimes with digital sound-alikes ===
Juho Kunsola (talk | contribs) (→Digital sound-alikes: various improvements) |
Juho Kunsola (talk | contribs) (→Digital sound-alikes: mv contents + change subheading to === Documented crimes with digital sound-alikes ===) |
||
| Line 116: | Line 116: | ||
Then in '''2018''' at the '''[[w:Conference on Neural Information Processing Systems]]''' (NeurIPS) the work [http://papers.nips.cc/paper/7700-transfer-learning-from-speaker-verification-to-multispeaker-text-to-speech-synthesis 'Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis'] ([https://arxiv.org/abs/1806.04558 at arXiv.org]) was presented. The pre-trained model is able to steal voices from a sample of only '''5 seconds''' with almost convincing results | Then in '''2018''' at the '''[[w:Conference on Neural Information Processing Systems]]''' (NeurIPS) the work [http://papers.nips.cc/paper/7700-transfer-learning-from-speaker-verification-to-multispeaker-text-to-speech-synthesis 'Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis'] ([https://arxiv.org/abs/1806.04558 at arXiv.org]) was presented. The pre-trained model is able to steal voices from a sample of only '''5 seconds''' with almost convincing results | ||
The Iframe below is transcluded from [https://google.github.io/tacotron/publications/speaker_adaptation/ 'Audio samples from "Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis"' at google.gituhub.io], the audio samples of a sound-like-anyone machine presented as at the 2018 [[w:NeurIPS]] conference by Google researchers. | |||
{{#Widget:Iframe - Audio samples from Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis by Google Research}} | {{#Widget:Iframe - Audio samples from Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis by Google Research}} | ||
Observe how good the "VCTK p240" system is at deceiving to think that it is a person that is doing the talking. | |||
<section end=GoogleTransferLearning2018 /> | <section end=GoogleTransferLearning2018 /> | ||
| Line 127: | Line 128: | ||
{{#ev:youtube|0sR1rU3gLzQ|640px|right|Video [https://www.youtube.com/watch?v=0sR1rU3gLzQ video 'This AI Clones Your Voice After Listening for 5 Seconds' by '2 minute papers' at YouTube] describes the voice thieving machine by Google Research in [[w:NeurIPS|w:NeurIPS]] 2018.}} | {{#ev:youtube|0sR1rU3gLzQ|640px|right|Video [https://www.youtube.com/watch?v=0sR1rU3gLzQ video 'This AI Clones Your Voice After Listening for 5 Seconds' by '2 minute papers' at YouTube] describes the voice thieving machine by Google Research in [[w:NeurIPS|w:NeurIPS]] 2018.}} | ||
=== | === Documented crimes with digital sound-alikes === | ||
In 2019 reports of crimes being committed with digital sound-alikes started surfacing. As of Jan 2022 no reports of other types of attack than fraud have been found. | In 2019 reports of crimes being committed with digital sound-alikes started surfacing. As of Jan 2022 no reports of other types of attack than fraud have been found. | ||
| Line 169: | Line 163: | ||
** [https://www.unite.ai/deepfaked-voice-enabled-35-million-bank-heist-in-2020/ '''''Deepfaked Voice Enabled $35 Million Bank Heist in 2020''''' at unite.ai]<ref group="1st seen in">https://www.reddit.com/r/VocalSynthesis/</ref> reporting updated on 2021-10-15 | ** [https://www.unite.ai/deepfaked-voice-enabled-35-million-bank-heist-in-2020/ '''''Deepfaked Voice Enabled $35 Million Bank Heist in 2020''''' at unite.ai]<ref group="1st seen in">https://www.reddit.com/r/VocalSynthesis/</ref> reporting updated on 2021-10-15 | ||
<section end=2021 digital sound-alike enabled fraud /> | <section end=2021 digital sound-alike enabled fraud /> | ||
=== What should we do about digital sound-alikes? === | |||
Living people can defend<ref group="footnote" name="judiciary maybe not aware">Whether a suspect can defend against faked synthetic speech that sounds like him/her depends on how up-to-date the judiciary is. If no information and instructions about digital sound-alikes have been given to the judiciary, they likely will not believe the defense of denying that the recording is of the suspect's voice.</ref> themselves against digital sound-alike by denying the things the digital sound-alike says if they are presented to the target, but dead people cannot. Digital sound-alikes offer criminals new disinformation attack vectors and wreak havoc on provability. | |||
For these reasons the bannable '''raw materials''' i.e. covert voice models '''[[Law proposals to ban covert modeling|should be prohibited by law]]''' in order to protect humans from abuse by criminal parties. | |||
=== Example of a hypothetical 4-victim digital sound-alike attack === | === Example of a hypothetical 4-victim digital sound-alike attack === | ||