3 ms·
Privacy, where there's a problem if some original data can be inferred/made out (even using a white box attack against the model), is a higher bar than whether
by Ukv 3y ago
Privacy, where there's a problem if some original data can be inferred/made out (even using a white box attack against the model), is a higher bar than whether an image generator avoids copyright-infringing output under non-adversarial usage. Additionally, compared to image data, text is more prone to exact matches due to lower dimensionality and usually training with less data per parameter.
While it's still a topic deserving of research and mitigation, by the time your information has been scooped up by Common Crawl and trained on by some LLM it's probably in many other places that attackers are more realistically likely to look (search engine caches, Common Crawl downloads, sites specifically for scooping credentials, ...) before trying to extract it from the LLM.
- Retric 3y agoThe privacy issue isn’t just about the data being available as people’s names, addresses, and phone numbers are generally available. The issue is if they show up as part of some meme chat and then you as the LLM creator get sued because people start harassing them. In terms of copyright infringement the bar is quite low, and copying is a basic part of how these algorithms work. This may or may not be an issue for you personally but it is a large land mine for commercial use especially if you’re independently creating one of these systems.
- Ukv 3y ago> The issue is if they show up as part of some meme chat and then you as the LLM creator get sued because people start harassing them. This seems a more obscure concern than extraction of data. > copying is a basic part of how these algorithms work Do you mean during training/gradient descent, or reverse diffusion?