3,839
edits
Synthetic human-like fakes: Difference between revisions
→Handwriting syntheses: mentioning that here is similar risk to what happened in 2018 when w:speaker recognition knowledge was applied to a generative task with w:Transfer learning
Juho Kunsola (talk | contribs) (→Handwriting syntheses: + Recurrent neural network handwriting generation demo at cs.toronto.edu is a demonstration site for publication Generating Sequences With Recurrent Neural Networks at arxiv.org...) |
Juho Kunsola (talk | contribs) (→Handwriting syntheses: mentioning that here is similar risk to what happened in 2018 when w:speaker recognition knowledge was applied to a generative task with w:Transfer learning) |
||
| Line 245: | Line 245: | ||
# Defensively, to hide one's handwriting style from public view | # Defensively, to hide one's handwriting style from public view | ||
# Offensively, to thieve somebody else's handwriting style | # Offensively, to thieve somebody else's handwriting style | ||
Here we find a similar risk to that which realized when the [[w:speaker recognition]] systems turning out to be instrumental in the development of [[#Digital sound-alikes|digital sound-alikes]]. After the knowledge needed to recognize a speaker was [[w:Transfer learning|w:transferred]] into a generative task in 2018 by Google researchers, we no longer cannot effectively determine for English speakers which recording is human in origin and which is from a machine origin. | |||
Some syntheses: | Some syntheses: | ||