3 ms·

What's the practical difference between a synthetic fake profile photo and a stolen filtered real one?

by hopler 8y ago

AndrewKemendo 8y ago
Speed of content generation. You can create thousands of synth ones a second.
- gambler 8y ago
  Since it uses a deep neural network, I don't think it's "thousands a second". Also, you can download an image and crop a face out of it in several seconds. You could even automate the process. The biggest problem is transferring faces to existing photos. It was hard to do manually. Now it's much easier. Also, people are generally trained to ignore various artifacts by CGI-ridden movies and compression algorithms. So much of our notion of how the world looks comes from digital imagery, it's kind of scary. I think we need to change the threshold of quality for an image/video to constitute "proof" of any kind. You can hide most of the weird artifacts by scaling things down or passing them through heavy compression.
  - gwern 8y ago
    > Since it uses a deep neural network, I don't think it's "thousands a second". The generator is like 150MB. The forward pass is <0.1s. Hypothetically you should be able to generate on a decent GPU like a 1080ti with 11GB VRAM at full utilization <730 images per second. Use a few GPUs and you're at thousands per second.
    - gambler 8y ago
      Where is this data from? I see 300MB model on their Google Drive. And if I understand correctly, you also need source and destination images to transfer styles from and to, so it's not like the model generates photos out of thin air.
      - gwern 8y ago
        The 300MB model covers both the G and D. You only need G to generate. The style transfer is just noise. And I time my own 512px anime StyleGANs at ~21 images per second per model; half that throughput to account for the increased model size and depth of a 1024px. No matter how you tweak the numbers - halve it again if you wish! - it's clear that thousands per second is entirely attainable with a few GPUs at low cost. (For comparison, 8 V100s is ~$7/hr on AWS; 10x1080ti is ~$1.3/hr on Vast.ai.)
michaelbuckbee 8y ago
Stolen ones are easier to do reverse image detection on and expose as fraudulent. I see this a lot on Twitter - bots pull a mix of stolen profile pics, bios, etc.
kzrdude 8y ago
Possibly ethics and law (personal rights to the photo)?

What's the practical difference between a synthetic fake profile photo and a stolen filtered real one?

by hopler 8y ago

What's the practical difference between a synthetic fake profile photo and a stolen filtered real one?

AndrewKemendo 8y ago

Speed of content generation. You can create thousands of synth ones a second.

gambler 8y ago
Since it uses a deep neural network, I don't think it's "thousands a second". Also, you can download an image and crop a face out of it in several seconds. You could even automate the process. The biggest problem is transferring faces to existing photos. It was hard to do manually. Now it's much easier. Also, people are generally trained to ignore various artifacts by CGI-ridden movies and compression algorithms. So much of our notion of how the world looks comes from digital imagery, it's kind of scary. I think we need to change the threshold of quality for an image/video to constitute "proof" of any kind. You can hide most of the weird artifacts by scaling things down or passing them through heavy compression.
- gwern 8y ago
  > Since it uses a deep neural network, I don't think it's "thousands a second". The generator is like 150MB. The forward pass is <0.1s. Hypothetically you should be able to generate on a decent GPU like a 1080ti with 11GB VRAM at full utilization <730 images per second. Use a few GPUs and you're at thousands per second.
  - gambler 8y ago
    Where is this data from? I see 300MB model on their Google Drive. And if I understand correctly, you also need source and destination images to transfer styles from and to, so it's not like the model generates photos out of thin air.
    - gwern 8y ago
      The 300MB model covers both the G and D. You only need G to generate. The style transfer is just noise. And I time my own 512px anime StyleGANs at ~21 images per second per model; half that throughput to account for the increased model size and depth of a 1024px. No matter how you tweak the numbers - halve it again if you wish! - it's clear that thousands per second is entirely attainable with a few GPUs at low cost. (For comparison, 8 V100s is ~$7/hr on AWS; 10x1080ti is ~$1.3/hr on Vast.ai.)

michaelbuckbee 8y ago

Stolen ones are easier to do reverse image detection on and expose as fraudulent. I see this a lot on Twitter - bots pull a mix of stolen profile pics, bios, etc.

kzrdude 8y ago

Possibly ethics and law (personal rights to the photo)?