grun
Joined 15 years ago
Joined 15 years ago
- confidential computing and trusted execution environments (TEEs) are the strongest primitive yet invented to prevent model weight exfiltration. there are other techniques. like bandwidth limiting of GPU workers and key sealing. but none as strong as those afforded by confidential computing and TEEs
- see https://confidential.ai/docs/confidential-computing-primer for a primer on confidential computing
- using confidential computing and TEEs to protect model weights in use is an area of active research
note: i work on confidential.ai
if you also like keeping your hands on the keyboard, give it a try; it could click for you, too (pun intended)
For example, to intercept network requests to serve them from the peer-to-peer network, a <![CDATA[ tag had to be inserted into the <head> of the document to block rendering of the subsequent document. Then, once the document had finished downloading, the page HTML was manually rendered so all assets (eg JS, CSS, images, etc tags) could be loaded via JS instead of the browser natively. This was both slow and resulted in empty white pages on load. Now? We have the Service Worker API. (https://developer.mozilla.org/en-US/docs/Web/API/Service_Wor...)
I'm not as familiar with Peer5's tech stack, so I can't speak there. But hi Shachar!
> How does the website possibly detect that I'm on a cellular network when I'm a) use Firefox (which doesn't implement the NetworkInformation API) and b) use a VPN to either a data center or my home server?
Yep. An IP lookup is also done to see which AS the user is on. Eg an AWS IP vs a T-Mobile IP.
We also do this to detect when people tether, as when tethering Chrome will report a Wi-Fi connection via the NetworkInforamtion API but it's Wi-Fi on top of an underlying cellular connection.
> There's also the massive privacy issue: other people will know what websites you visit by simply using the P2P system
All cached data is both fragmented and encrypted. When a node sends data to another peer, it's an encrypted fragment of a file and the sender doesn't know 1) what data it's sending nor 2) which website that data is being sent for.
All content is fragmented, encrypted, and hashed before it's distributed across the network. If a peer ever receives a file piece from a peer and that piece's hash doesn't match the expected hash, it's dropped along with the connection to that peer.