4 ms·

Is there a way to see if images on my website are in a training data set?

by cactusplant7374 2y ago

_heimdall 2y ago
If the site is public, you content is in all the major training sets.
- bee_rider 2y ago
  I wonder how long until something like Copilot+, but also with the ability to leak these images into the training data, is released by somebody (probably not by MS, they appear to be trying to maintain some appearance of not being entirely malware authors). I wonder if “affirming the consequent” is still ok (not that you’ve done so, your post just brought it to mind).
  - _heimdall 2y ago
    Affirming the consequent is always an interesting one to me when it overlaps with a tragedy of the commons situation. If publicly available content can be used in training sets, any one model willing to use them will have an advantage. Is it reasonable to assume then that all publicly available content is in the most popular training sets, or is that falling into the affirming the consequent fallacy?
mrinfinitiesx 2y ago
Assume that everything public facing is
- HeatrayEnjoyer 2y ago
  It's worse. Rumor has it OpenAI pays blackhat groups for access to semi-private and private datasets. e.g. Discord servers, corporate O365 inboxes.
  - mrinfinitiesx 2y ago
    thank you for the headsup
  - ronsor 2y ago
    I can believe Discord servers, but not corporate inboxes.
    - HeatrayEnjoyer 2y ago
      It wouldn't be the most outrageous or illegal thing they have done. In 2022 a Kenyan data labeling firm suddenly cut ties mid-contract because OpenAI was asking them to gather and label CSAM images.[0] >In February 2022, Sama and OpenAI’s relationship briefly deepened, only to falter. That month, Sama began pilot work for a separate project for OpenAI: collecting sexual and violent images—some of them illegal under U.S. law—to deliver to OpenAI. The work of labeling images appears to be unrelated to ChatGPT. In a statement, an OpenAI spokesperson did not specify the purpose of the images the company sought from Sama, but said labeling harmful images was “a necessary step” in making its AI tools safer. (OpenAI also builds image-generation technology.) In February, according to one billing document reviewed by TIME, Sama delivered OpenAI a sample batch of 1,400 images. Some of those images were categorized as “C4”—OpenAI’s internal label denoting child sexual abuse—according to the document. Also included in the batch were “C3” images (including bestiality, rape, and sexual slavery,) and “V3” images depicting graphic detail of death, violence or serious physical injury, according to the billing document. OpenAI paid Sama a total of $787.50 for collecting the images, the document shows. >Within weeks, Sama had canceled all its work for OpenAI—eight months earlier than agreed in the contracts. The outsourcing company said in a statement that its agreement to collect images for OpenAI did not include any reference to illegal content, and it was only after the work had begun that OpenAI sent “additional instructions” referring to “some illegal categories.” “The East Africa team raised concerns to our executives right away. Sama immediately ended the image classification pilot and gave notice that we would cancel all remaining [projects] with OpenAI,” a Sama spokesperson said. [0] https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
      - 2y ago
KiranRao0 2y ago
You have to assume all publicly available data is used for training. If you don’t want people using your images, consider data poisoning techniques to punish systems that train on them. https://www.lakera.ai/blog/training-data-poisoning https://www.lakera.ai/blog/training-data-poisoning https://www.technologyreview.com/2023/10/23/1082189/data-poisoning-artists-fight-generative-ai/ https://www.technologyreview.com/2023/10/23/1082189/data-poi...
- ronsor 2y ago
  These tools don't actually work[0], by the way. [0] https://arxiv.org/abs/2406.12027 https://arxiv.org/abs/2406.12027
  - KiranRao0 2y ago
    Very interesting article. I'm guessing this will remain a cat/mouse game for a while. However, any moves will increase the cost of training on un-licensed internet data. This will shift the balance slightly closer towards AI companies licensing data as opposed to using scraped data. Emphasis on slightly.
- jimmaswell 2y ago
  What a senseless, selfish, petty thing to attempt. To not only deprive systems ultimately benefitting humanity of training content for no possible tangible benefit to yourself, but to try to sabotage them because you think it's theft for a computer to learn from patterns in an image no differently from a human artist (who will often outright trace existing works to train their neural nets, but it's ok when they do it because uh.. uh humans are special!)
  - throw10920 2y ago
    > What a senseless, selfish, petty thing to attempt This is clearly not selfishness. Selfishness is believing you're entitled to other people's work for free, which is the implicit assumption made by people claiming that they should be able to train their models on your work. Selfishness is trying to take other people's work and turn it onto an AI model without compensating them on their turns. This demonstrates extreme entitlement, lack of a consistent moral system, and a fundamental ignorance of basic economics. > no differently from a human artist Yes, humans are objectively different from computers, including neural networks. Humans are sentient, and existing AI is not. It's not even accurate to say that humans have "neural nets" in their brains, because human cognition not understood. The AI poisoning technique is very moral - if someone doesn't explicitly contact me and ask me for permission to train on my data, then they absolutely deserve a corrupt model. I think I'll do this.
Zenzero 2y ago
There's nothing stopping anyone here from coming to your public website and downloading your images for use in a training set. If someone did that, why would they be posting online where they pulled data from?

Is there a way to see if images on my website are in a training data set?

_heimdall 2y ago

If the site is public, you content is in all the major training sets.

bee_rider 2y ago
I wonder how long until something like Copilot+, but also with the ability to leak these images into the training data, is released by somebody (probably not by MS, they appear to be trying to maintain some appearance of not being entirely malware authors). I wonder if “affirming the consequent” is still ok (not that you’ve done so, your post just brought it to mind).
- _heimdall 2y ago
  Affirming the consequent is always an interesting one to me when it overlaps with a tragedy of the commons situation. If publicly available content can be used in training sets, any one model willing to use them will have an advantage. Is it reasonable to assume then that all publicly available content is in the most popular training sets, or is that falling into the affirming the consequent fallacy?

mrinfinitiesx 2y ago

Assume that everything public facing is

HeatrayEnjoyer 2y ago
It's worse. Rumor has it OpenAI pays blackhat groups for access to semi-private and private datasets. e.g. Discord servers, corporate O365 inboxes.
- mrinfinitiesx 2y ago
  thank you for the headsup
- ronsor 2y ago
  I can believe Discord servers, but not corporate inboxes.
  - HeatrayEnjoyer 2y ago
    It wouldn't be the most outrageous or illegal thing they have done. In 2022 a Kenyan data labeling firm suddenly cut ties mid-contract because OpenAI was asking them to gather and label CSAM images.[0] >In February 2022, Sama and OpenAI’s relationship briefly deepened, only to falter. That month, Sama began pilot work for a separate project for OpenAI: collecting sexual and violent images—some of them illegal under U.S. law—to deliver to OpenAI. The work of labeling images appears to be unrelated to ChatGPT. In a statement, an OpenAI spokesperson did not specify the purpose of the images the company sought from Sama, but said labeling harmful images was “a necessary step” in making its AI tools safer. (OpenAI also builds image-generation technology.) In February, according to one billing document reviewed by TIME, Sama delivered OpenAI a sample batch of 1,400 images. Some of those images were categorized as “C4”—OpenAI’s internal label denoting child sexual abuse—according to the document. Also included in the batch were “C3” images (including bestiality, rape, and sexual slavery,) and “V3” images depicting graphic detail of death, violence or serious physical injury, according to the billing document. OpenAI paid Sama a total of $787.50 for collecting the images, the document shows. >Within weeks, Sama had canceled all its work for OpenAI—eight months earlier than agreed in the contracts. The outsourcing company said in a statement that its agreement to collect images for OpenAI did not include any reference to illegal content, and it was only after the work had begun that OpenAI sent “additional instructions” referring to “some illegal categories.” “The East Africa team raised concerns to our executives right away. Sama immediately ended the image classification pilot and gave notice that we would cancel all remaining [projects] with OpenAI,” a Sama spokesperson said. [0] https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
    - 2y ago

KiranRao0 2y ago

You have to assume all publicly available data is used for training. If you don’t want people using your images, consider data poisoning techniques to punish systems that train on them. https://www.lakera.ai/blog/training-data-poisoning https://www.lakera.ai/blog/training-data-poisoning https://www.technologyreview.com/2023/10/23/1082189/data-poisoning-artists-fight-generative-ai/ https://www.technologyreview.com/2023/10/23/1082189/data-poi...

ronsor 2y ago
These tools don't actually work[0], by the way. [0] https://arxiv.org/abs/2406.12027 https://arxiv.org/abs/2406.12027
- KiranRao0 2y ago
  Very interesting article. I'm guessing this will remain a cat/mouse game for a while. However, any moves will increase the cost of training on un-licensed internet data. This will shift the balance slightly closer towards AI companies licensing data as opposed to using scraped data. Emphasis on slightly.
jimmaswell 2y ago
What a senseless, selfish, petty thing to attempt. To not only deprive systems ultimately benefitting humanity of training content for no possible tangible benefit to yourself, but to try to sabotage them because you think it's theft for a computer to learn from patterns in an image no differently from a human artist (who will often outright trace existing works to train their neural nets, but it's ok when they do it because uh.. uh humans are special!)
- throw10920 2y ago
  > What a senseless, selfish, petty thing to attempt This is clearly not selfishness. Selfishness is believing you're entitled to other people's work for free, which is the implicit assumption made by people claiming that they should be able to train their models on your work. Selfishness is trying to take other people's work and turn it onto an AI model without compensating them on their turns. This demonstrates extreme entitlement, lack of a consistent moral system, and a fundamental ignorance of basic economics. > no differently from a human artist Yes, humans are objectively different from computers, including neural networks. Humans are sentient, and existing AI is not. It's not even accurate to say that humans have "neural nets" in their brains, because human cognition not understood. The AI poisoning technique is very moral - if someone doesn't explicitly contact me and ask me for permission to train on my data, then they absolutely deserve a corrupt model. I think I'll do this.

Zenzero 2y ago

There's nothing stopping anyone here from coming to your public website and downloading your images for use in a training set. If someone did that, why would they be posting online where they pulled data from?