Hacker News Reader

Comment by blauditore | Hacker News Reader

I'm trying to understand what kind of computations would be done on encrypted data. For cases where computation depends on the raw data (conditionally), this would not be feasible. For other cases, such as e.g. training some ML model to map inputs to output, this could also be done on any non-homomophically ecrypted data. What are actually use cases for computations on FHE data?

Edit: Thinking of it more, I could imagine some sort of image transformations (e.g. a blur filter) as applications - but those are fairly corner use cases, are there more broad ones?

View on HN

Potentially do stats on website usage data without ever seeing the actual data (the endgoal is not super humanitarian I know). Medical diagnosis could be one, bonus is the service never sees your data.

Image filtering made me think of this FHE demo/quickstart: https://6min.zama.ai/

As per conditional programs, there might be a way to do it but it would always be the worst case (no early loop break for example) so the runtimes would be aweful and the memory state could get prohibitively large I guess (to keep track of all the branches).

Medical diagnosis is a great example!

HE allows third-parties to perform processing without seeing the plaintext. This allows sensitive data to be held and processed by a cloud computing platform. It could also be useful for untrusted peer-to-peer computing (something like Ethereum). It seems too inefficient to be practical at the moment, but on the other hand that hasn't stopped cryptocurrencies before...

Hospital saves data about a patient in a theoretical common file that tracks all his health info, but solely belongs to the patient. Insurance wants to add data about some reimbursements for glasses they've done, but should not be allowed to see his medical conditions.

FHE lets you work on the data and update it without ever revealing the contents.

So, you could do the same with an append-only log with no read permissions?

Accept that, in the case of your append-only log, you need to trust that someone is applying that lack of read permission to themselves. And that they do it well.

Couldn't regular asymmetric encryption take care of that? Just encrypt each entry with the patient's public key.

I think you only need homomorphic encryption if the transformation is something that the owner of data can't or shouldn't do themselves. It would be overkill for simple financial information.

Absolutely. That’s why gp’s parent mentioned working on the patients’ data, not just updating it.

For example, running machine learning models on homomorphically encrypted health data.

*Except that

> are there more broad ones?

Indeed there are. FHE could provide a way to stay compliant with privacy regulations while outsourcing processing to external providers.

Think medical data or trade secrets.

Reply on news.ycombinator.com