Confidential computing and AI fit together | Hacker News Reader

Confidential computing and AI fit together

edgeless.systems

51 pointstranspute5 comments

One use case that wasn't mentioned but is arguably the most important of all:

Keeping peoples' data protected from data hungry tools and processes, to protect people from penetrating queries that could leave them vulnerable and globally exposed.

The queriability of unencrypted data is just gonna keep increasing from here on out.

Cryptography is the only way to keep the digital world from collapsing in on itself. If there is to be any sense of locality, separation or closeness then cryptography is pivotal.

Cryptography wards off the "tower of babel" qualities of this new tech.

In the current Confidential Computing landscape, trust is being shifted from the cloud service provider to the hardware vendor. Whether it's AMD-SEV or Intel's SGX/TDX, the trust "root" becomes the entity responsible for creating the secure enclave or confidential VM.

It's important to recognize that this 'root of trust' must still exist somewhere.

While homomorphic encryption offers an possibility for computational privacy without the need for this kind of trust, it isn't a one-size-fits-all solution due to its computational overhead and limitations. (although I'm not an expert on that).

Although that's partly true, bear in mind that not all RoTs are the same. Having a hardware vendor as the RoT is better because the manufacturer has no good way to abuse their position: you can easily firewall the hardware off from the chip vendors, or otherwise monitor that it's not phoning home, and hardware vendors have no straightforward way to access their customers machines even in the absence of CC. This isn't a meaningful thing to attempt for a cloud vendor.

In most places, running the model isn't actually the real problem for security, its keeping the training set secure.

More over, accidentally leaking the dataset through crafted prompts.

Its great to see some level of verifiability of hardware, which is an improvement.

Most exciting aspect of confidential computing is it’s inevitable marriage with blockchain. Imagine being able to monetize “secrets” accessible only to programs running in TEEs. It’s probably not going to amount to much but iExec in France is trying to make an economy of secrets secured by confidential computing and are giving out boku bucks for devs/bounties/grants

Reply on news.ycombinator.com