Fully homomorphic encryption on a chip?

robertmcgrath

Dec 27

One response to the coming quantum decryption disaster is new encryption methods that don't rely on problems that can be cracked by quantum computers. A leading candidate are fully homomorphic encryption (FHE)schemes, which I surely don't understand in any non-trivial way.

In addition to being hard to crack, these schemes have the super- mega- ultra-cool feature that it is possible to compute on the encrypted data. I.e., it isn't necessary to decrypt the data in order to use it. This is especially useful in networked systems where you need to pass around data, and compute on remote nodes. The data remains safely encrypted everywhere.

Magic.

Unsurprisingly, this magic has a price, in that the computation is generally slower using encrypted data. Possibly a lot slower. Possibly too slow to be useful. As in a calculation that takes a fraction of a second might take a week. (!)

But, wait.

Samuel K. Moore reports that, in his words, "Chips to Compute With Encrypted Data Are Coming" [1]. Specialized chips are being developed that will do the FHE operations really fast. These chips basically only work with encrypted data, so not only can the data be secure, it has to be secure.

I gather that the basic idea is that the encrypted data represents the data as really big numbers. Huge numbers. So the specialized chips are designed to handle this data effectively. There are multiple ways to skin this cat, and several are coming near production.

Neat.

I'll note that these chips in themselves are not the complete solution. For one thing, encrypting the data in the first place is a tough engineering problem, with humans in the loop to mess things up. For another, even the completely encrypted computations may leak information (e.g., the activity itself may reveal information). And, let's not forget that encryption algorithms can have bugs.

The bottom line is, these chips might work perfectly, but if the encryption itself is cracked, or there is a bug somewhere in the system, or any other security failure, then the fancy hardware will mean nothing.

Sound like a fun area to work in, no?

Samuel K. Moore, Chips to Compute With Encrypted Data Are Coming in IEEE Spectrum - Semiconductors, December 22 2023. https://spectrum.ieee.org/homomorphic-encryption