Dave Archer

CTO, Niobium Microsystems | Principal Scientist, Galois, Inc.

[Dave Archer @DaveArcher37468]

We have the math to check this stuff without seeing the underlying data. At some point “we need to store it” stops being a technical argument.

[Niobium @NiobiumInc]

153M driver's license scans are for sale on the dark web, likely leaked by an ID verification vendor.

A license check only needs to confirm three things: real, valid, and of age. The vendor stored the entire document.

On the Fog™, that archive never exists. Steal everything we hold, and you get ciphertext.

https://t.co/ayByigYozK

[Niobium @NiobiumInc]

153M driver's license scans are for sale on the dark web, likely leaked by an ID verification vendor.

A license check only needs to confirm three things: real, valid, and of age. The vendor stored the entire document.

On the Fog™, that archive never exists. Steal everything we hold, and you get ciphertext.

https://t.co/ayByigYozK

[Dave Archer @DaveArcher37468]

@T3chFalcon Wearables already sit on an unusually intimate stream of personal data. Add ambient audio and the privacy problem expands from “my data” to “everyone around me.”

[Leila @Leiladb22]

OpenAI after 10,000 agents “find” the answer:

[Leila @Leiladb22]

When the homework is already in the group chat

[Dave Archer @DaveArcher37468]

@edzitron The models are too useful for people to stop using them. Sensitive inference just needs to move under FHE.

[Dave Archer @DaveArcher37468]

Feels increasingly obvious that sensitive AI inference is going to happen under FHE.

Use the model. Keep your data encrypted. The provider never gets the plaintext.

If you need a use case, ask those two Navier-Stokes guys🤷

[Dave Archer @DaveArcher37468]

@wildtypehuman There’s a third option: don’t give the model provider the plaintext in the first place. This is exactly where fully homomorphic encryption is headed.

[Dave Archer @DaveArcher37468]

@Exomnium OpenAI right now: who? Me?

[Dave Archer @DaveArcher37468]

@ns123abc “We cannot rule out” is what my vet says about tumors, not what a company should say about your data

[Dave Archer @DaveArcher37468]

“De-identified data” from a pool of, generously, two people solving a Millennium Prize problem. Incredible.

OR we could just run inference under FHE and never give the provider the plaintext to begin with🤷

[The All-In Podcast @theallinpod]

Chamath Warns: Your AI Data Isn’t as Private as You Think

Today's statement from @OpenAI on the Navier-Stokes Problem and whether it used recent user data to boost its own capabilities:

“While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”

Here's what @chamath had to say on AI's fragile privacy and the ZDR myth back in July:

“Privacy in AI is very fragile and it's very brittle. There are all kinds of non-obvious data leak vectors lurking in AI.

If you think that you're going to flip a ZDR switch, zero data retention, which is the magic term that the industry uses to tell you that everything's going to be okay.

I think the answer and the message should be, ‘It's not going to be okay because you can't guarantee any of it.’

So the model companies, when they give you these zero data retention policies, are probably trying their best.

But I think the reality is you are leaking information where you don't know it. And they, despite their best efforts, may still have trapdoors that they don't even know about until it's figured out by somebody else.

You need an independent third-party layer to interface to these models to manage this exposure, because there are trapdoors everywhere.”

[Niobium @NiobiumInc]

🚨 New #TGIFHE

For OpenAI Build Week, our MBA intern Sarah built StealthMatch, an FHE demo, using Codex.

Founders rank investors and grants. Raise targets, runway, and conflict lists stay encrypted with CKKS the whole time.

Runs on CPU, a simulator, or Niobium FPGAs via The Fog™.
https://t.co/9ukaPz4zwX

[Niobium @NiobiumInc]

Most people think encryption works because nobody knows how it works.

On the contrary, all of the algorithms and research are public. The security lives in the math.

Our latest FHE 101 post talks through how that math works.

https://t.co/OrkeG7nslx

[Niobium @NiobiumInc]

A preliminary draft went up on the IACR ePrint archive last week claiming a quantum algorithm that could break the lattice-based math underlying post-quantum encryption and every practical FHE scheme, including the ones our hardware accelerates, no matter how large the keys.

Four days later, Aparna Gupte, Seyoon Ragavan and Mark Zhandry published a proof that it can't recover even one bit of a secret key better than a random guess.

Our CTO @DaveArcher37468 on what happened:
https://t.co/Mq4oQbNR5Z

[Dave Archer @DaveArcher37468]

A paper appeared last week that, if correct, would have broken the NIST post-quantum standards and every practical FHE scheme. It wasn't correct.
The whole thing took four days. Published openly, checked by people who knew exactly where to look, and resolved in public.

I wrote about it here: https://t.co/kz0hyhBDhT

[Dave Archer @DaveArcher37468]

Great to see HEIR getting this kind of attention. More to come

[Niobium @NiobiumInc]

Google Security featured four private inference applications in its new post on HEIR.

We worked on two of them.
Credit card fraud detection with Hardshell, and anomaly detection over encrypted network traffic using Kitsune.

Glad to be building alongside Google as more FHE makes its way out of the lab.

https://t.co/JUgLQtHfDQ

[Dave Archer @DaveArcher37468]

@NiobiumInc A good reminder that you don’t need to understand every cryptographic detail before you can start building useful things with FHE.

[Dave Archer @DaveArcher37468]

A nice FHE demo from @FlightlessBrad: Offmeter.

A household encrypts a 24-hour smart-meter profile, a server evaluates an eligibility model directly on the ciphertext, and the encrypted result comes back for client-side decryption.

What I like about this one is that the privacy boundary is easy to inspect. The secret key stays client-side, and the demo explicitly refuses to run if that key appears on the server.

If you want to see what “compute on encrypted data” actually looks like without a UI hiding the mechanics, the README walks through the whole thing: https://t.co/Qjbjq7a1Ha

[Dave Archer @DaveArcher37468]

Most people probably don’t realize how much data gets collected on them just from leaving the house. You should be able to go for a run without having your face and movements tracked across multiple cameras and stored in systems you know nothing about. Excuse the hot take, but that should probably bother us way more than it does

[Dave Archer @DaveArcher37468]

@Leiladb22 @NiobiumInc The bar for "FHE is too hard" just moved