How do human brains detect false irregularities in faces?

[ally]

Hi!

So my question is: how to human brains detect false irregularities in faces? With false irregularities in faces I mean aspects of for example deepfake faces or photoshopped faces which show irregularities that are not normal to appear in an observed human face (for example weird lines around eyebrows in deepfakes or colour irregularities in the face).

Are there specific brain regions that are used when such irregularities are observed by the human during facial perception? Maybe someone can send a link to relevant papers (I'm really struggling to find them myself)?

Thanks!

 

[phinds]

I think these days, if you can tell it's an AI generated face, it's amateur hour in graphics-land. That may not be true for deepfakes of people you recognize, like movie stars, but given nothing more than a picture with no other knowledge of who the picture is of I doubt you could tell that a good AI generated image is not a real person.

 

[jim mcnamara]

There is literature on deepfake identification using AI or other methods:
alternate method with neural network technology
"Detecting Deepfake-Forged Contents with Separable Convolutional Neural Network and Image Segmentation"
https://arxiv.org/abs/1912.12184

I do not see much on how humans detect them, or even if detection is 100%.

The point being you may be able to 'reverse engineer' one of the alternate methods back to a paper that deals with humans. Pubmed frequently has a link to who referenced the current paper. Impact is another way.

On a more superficial level wikipedia has links to get you started, example:
https://en.wikipedia.org/wiki/Deepfake
leads to:
https://news.berkeley.edu/2019/06/18/researchers-use-facial-quirks-to-unmask-deepfakes/

Someone else may have more direct knowledge - best I can do.

 

[ally]

There is literature on deepfake identification using AI or other methods:
alternate method with neural network technology
"Detecting Deepfake-Forged Contents with Separable Convolutional Neural Network and Image Segmentation"
https://arxiv.org/abs/1912.12184

I do not see much on how humans detect them, or even if detection is 100%.

The point being you may be able to 'reverse engineer' one of the alternate methods back to a paper that deals with humans. Pubmed frequently has a link to who referenced the current paper. Impact is another way.

On a more superficial level wikipedia has links to get you started, example:
https://en.wikipedia.org/wiki/Deepfake
leads to:
https://news.berkeley.edu/2019/06/18/researchers-use-facial-quirks-to-unmask-deepfakes/

Someone else may have more direct knowledge - best I can do.

Thanks for the response :).

I'm doing my bachelor thesis in Artificial Intelligence about deepfake detection so I know about how deepfakes are being created and being detected artificially currently. But I'm curious about the specific brain regions humans use to detect such irregularities like how artificial methods for example look at the difference in illuminants between face centre and the face area around it. I assume it takes place in the fusiform face area, but can't find any specific literature about how humans distinguish those visible false irregularities from normal human faces. Maybe someone has an idea?