[HN Gopher] Let's talk about AI and end-to-end encryption
       ___________________________________________________________________
        
       Let's talk about AI and end-to-end encryption
        
       Author : chmaynard
       Score  : 134 points
       Date   : 2025-01-17 05:50 UTC (17 hours ago)
        
 (HTM) web link (blog.cryptographyengineering.com)
 (TXT) w3m dump (blog.cryptographyengineering.com)
        
       | crackalamoo wrote:
       | See also CrypTen, Meta's library for privacy preserving machine
       | learning: https://github.com/facebookresearch/CrypTen. This isn't
       | fully homomorphic encryption, but it is multi-party computation
       | (MPC), which hides the inputs from the company owning the model.
       | 
       | But while not revealing user input, it would still reveal the
       | outputs of the model to the company. And yeah, as the article
       | mentions, unfortunately this kind of thing (MPC or fully-
       | homomorphic encryption) probably won't be feasible for the most
       | powerful ML models.
        
       | EGreg wrote:
       | I heard that homomorphic encryption can actually preserve all the
       | operations in neural networks, since they are differentiable. Is
       | this true? What is the slowdown in practice?
        
         | crackalamoo wrote:
         | This is true in principle, yes. In practice, the way this
         | usually works is by converting inputs to bits and bytes, and
         | then computing the result as a digital circuit (AND, OR, XOR).
         | 
         | Doing this encrypted is very slow: without hardware
         | acceleration or special tricks, running the circuit is 1
         | million times slower than unencrypted, or about 1ms for a
         | single gate. (https://www.jeremykun.com/2024/05/04/fhe-
         | overview/)
         | 
         | When you think about all the individual logic gates involved in
         | just a matrix multiplication, and scale it up to a diffusion
         | model or large transformer, it gets infeasible very quickly.
        
           | j2kun wrote:
           | There are FHE schemes that do better than binary gates (cf.
           | CKKS) but they have other problems in that they require
           | polynomial approximations for all the activation functions.
           | Still they are much better than the binary-FHE schemes for
           | stuff like neural networks, and most hardware accelerators in
           | the pipeline right now are targeting CKKS and similar for
           | this reason.
           | 
           | For some numbers, a ResNet-20 inference can be done in CKKS
           | in like 5 minutes on CPU. With custom changes to the
           | architecture you can get less than one minute, and in my view
           | HW acceleration will improve that by another factor of 10-100
           | at least, so I'd expect 1s inference of these (still small)
           | networks within the next year or two.
           | 
           | LLMs, however, are still going to be unreasonably slow for a
           | long time.
        
       | jFriedensreich wrote:
       | I think this has also a silver lining. The E2E encryption
       | movement especially for messenger apps was largely also used to
       | silently lock users out of their own data and effectively prevent
       | user agency to use their own data to move apps, write automations
       | or archive, this is not just true for whatsapp (the data export
       | feature does not fully work since its launch and was just made to
       | appease some EU law that did not properly check if the button
       | works until the end.) Also signal does not have a way to do this.
       | Maybe with ai coming into the game companies finally decide to
       | provide access to data, I just hope it's in a transparent way
       | with user opt in and user control.
        
         | do_not_redeem wrote:
         | 1. Is data encrypted in transit?
         | 
         | 2. Can the user access their data at rest?
         | 
         | Those two things are entirely orthogonal.
         | 
         | I don't think you can extrapolate a trend from a few apps
         | having bugs in their export code. Google Takeout is also
         | notoriously buggy and they don't use E2E encryption. A more
         | likely explanation is companies of all kinds don't care that
         | much about export functionality, due to the incentives
         | involved.
        
           | jFriedensreich wrote:
           | you CAN extrapolate from nearly all e2e encrypted apps not
           | giving a way to use the data. And there is a big difference
           | between buggy google export features or facebook actively
           | making export unusable to lock in users.
        
         | walrus01 wrote:
         | Signal does not have a way to manually export your private keys
         | and chat history, but the process of "moving" your signal
         | account to a new phone is quite straightforward. You put both
         | devices on the same wifi/LAN layer 2 broadcast segment, start
         | the transfer process in the app, input the verification codes
         | displayed on the screen from both devices, and it sends
         | everything over. This moves the private key in a way that does
         | not result in all of your contacts receiving the scary "this
         | person's key has changed" message.
        
           | jFriedensreich wrote:
           | "Moving your account" is not what i talk about, besides not
           | being possible between android and ios with history. User
           | agency means a user is allowed to access his data and do what
           | they want with it how they want, realtime and with whatever
           | code they want to write to do so.
        
       | bobbiechen wrote:
       | There is always going to be a gap between local-first processing
       | and what can be achieved in a full-sized datacenter/cloud. That
       | leads to the risks mentioned in the article.
       | 
       | I wrote about Apple's Private Cloud Compute last year; for the
       | foreseeable future, I still think server-side Confidential
       | Computing is the most practical way to do processing without huge
       | privacy risks: https://www.anjuna.io/blog/apple-is-using-secure-
       | enclaves-to...
        
       | tonygiorgio wrote:
       | > Although PCC is currently unique to Apple, we can hope that
       | other privacy-focused services will soon crib the idea.
       | 
       | IMHO, Apple's PCC is a step in the right direction in terms of
       | general AI privacy nightmares where they are at today. It's not a
       | perfect system, since it's not fully transparent and auditable,
       | and I do not like their new opt-out photo scanning feature
       | running on PCC, but there really is a lot to be inspired by it.
       | 
       | My startup is going down this path ourselves, building on top of
       | AWS Nitro and Nvidia Confidential Compute to provide end to end
       | encryption from the AI user to the model running on the enclave
       | side of an H100. It's not very widely known that you can do this
       | with H100s but I really want to see this more in the next few
       | years.
        
         | mnahkies wrote:
         | I didn't actually realize that AWS supported this, I thought
         | Azure was the only one offering it
         | (https://azure.microsoft.com/en-us/blog/azure-confidential-
         | co...)
         | 
         | Are you speaking of this functionality?
         | https://developer.nvidia.com/blog/confidential-computing-on-...
         | (and am I just failing to find the relevant AWS docs?)
        
           | tonygiorgio wrote:
           | Yes, you're correct on both, though I think Google Cloud
           | recently started supporting it as well. AWS will likely have
           | GPU enclave support with Trainium 2 soon (AFAIK, that feature
           | is not publicly offered yet but could be wrong).
           | 
           | We work with Edgeless Systems who manages the GPU enclave on
           | Azure that we speak to from our AWS Nitro instance. While not
           | ideal, the power of enclaves and the attestation verification
           | process, we at least know that we're not leaking privacy by
           | going with a third party GPU enclave provider.
        
         | blueblimp wrote:
         | And the most important thing about PCC in my opinion is not the
         | technical aspect (though that's nice) but that Apple views user
         | privacy as something good to be maximized, differing from the
         | view championed by OpenAI and Anthropic (and also adopted by
         | Google and virtually every other major LLM provider by this
         | point) that user interactions must be surveilled for "safety"
         | purposes. The lack of privacy isn't due to a technical
         | limitation--it's intended, and they often brag about it.
        
           | natch wrote:
           | Something good to be maximized within the constraints of the
           | systems they have to work within. But at some point with
           | enough compromises it becomes maximizing the perception of
           | privacy, not the reality. Promoting these academic techniques
           | may just be perception management on the part of Apple, if
           | the keys are not controlled solely by the user.
        
       | bee_rider wrote:
       | The author helpfully emphasized the interesting question at the
       | end
       | 
       | > This future worries me because it doesn't really matter what
       | technical choices we make around privacy. It does not matter if
       | your model is running locally, or if it uses trusted cloud
       | hardware -- once a sufficiently-powerful general-purpose agent
       | has been deployed on your phone, the only question that remains
       | is who is given access to talk to it. Will it be only you? Or
       | will we prioritize the government's interest in monitoring its
       | citizens over various fuddy-duddy notions of individual privacy.
       | 
       | I do think there are interesting policy questions there. I mean
       | it could hypothetically be mandated that the government must be
       | given access to the agent (in the sense that we and these
       | companies exist in jurisdictions that can pass arbitrary laws;
       | let's skip the boring and locale specific discussion of whether
       | you think your local government would pass such a law).
       | 
       | But, on a technical level--it seems like it ought to be possible
       | to run an agent locally, on a system with full disk encryption,
       | and not allow anyone who doesn't have access to the system to
       | talk with it, right? So on a technical level I don't see how this
       | is any different from where we were previously. I mean you could
       | also run a bunch of regex's from the 80's to find whether or not
       | somebody has, whatever, communist pamphlets on their computers.
       | 
       | There's always been a question of whether the government should
       | be able to demand access to your computer. I guess it is good to
       | keep in mind that if they are demanding access to an AI agent
       | that ran on your computer, they are basically asking for a lossy
       | record of your entire hard drive.
        
         | _boffin_ wrote:
         | Unreasonable search?
        
           | bee_rider wrote:
           | > (in the sense that we and these companies exist in
           | jurisdictions that can pass arbitrary laws; let's skip the
           | boring and locale specific discussion of whether you think
           | your local government would pass such a law)
           | 
           | Anyway the idea of what's a reasonable search in the US has
           | been whittled away to almost nothing, right? "The dog smelled
           | weed on your hard drive." - A cop, probably.
        
           | chgs wrote:
           | Boring locale specific discussion.
        
         | cryptonector wrote:
         | > The author helpfully emphasized the interesting question at
         | the end
         | 
         | We're already there. AI or not doesn't affect the fact that
         | smartphones gather, store, and transmit a great deal of
         | information about their users and their users' actions and
         | interests.
        
       | walrus01 wrote:
       | It's a _good thing_ that encrypted data at rest on your local
       | device is inaccessible to cloud based  "AI" tools. The problem is
       | that your average person will blithely click
       | "yes/accept/proceed/continue/I consent" on pop up dialogs in a
       | GUI and agree to just about any Terms of Service, including
       | decrypting your data before it's sent to some "cloud" based
       | service.
       | 
       | I see "AI" tools being used even more in the future to
       | permanently tie people to monthly recurring billing services for
       | things like icloud, microsoft's personal grade of office365,
       | google workspace, etc. You'll pay $15 a month forever, and the
       | amount of your data and dependency on the cloud based provider
       | will mean that you have no viable path to ever stop paying it
       | without significant disruption to your life.
        
       | Animats wrote:
       | > Who does your AI agent actually work for?
       | 
       | Yes. I made that point a few weeks ago. The legal concept of
       | principal and agent applies.
       | 
       | Running all content through an AI in the cloud to check for
       | crimethink[1] is becoming a reality. Currently proposed:
       | 
       | - "Child Sexual Abuse Material", which is a growing category that
       | now includes AI-generated images in the US and may soon extend to
       | Japanese animation.
       | 
       | - Threats against important individuals. This may be extended to
       | include what used to be considered political speech in the US.
       | 
       | - Threats against the government. Already illegal in many
       | countries. Bear in mind that Trump likes to accuse people of
       | "treason" for things other than making war against the United
       | States.
       | 
       | - "Grooming" of minors, which is vague enough to cover most
       | interactions.
       | 
       | - Discussing drugs, sex, guns, gay activity, etc. Variously
       | prohibited in some countries.
       | 
       | - Organizing protests or labor unions. Prohibited in China and
       | already searched for.
       | 
       | Note that talking around the issue or jargon won't evade
       | censorship. LLMs can deal with that. Run some ebonics or
       | leetspeak through an LLM and ask it to translate it to standard
       | English. Translation will succeed. The LLM has probably seen more
       | of that dialect than most people.
       | 
       |  _" If you want a vision of the future, imagine a boot stepping
       | on a face, forever"_ - Orwell
       | 
       | [1] https://www.orwell.org/dictionary/
        
         | iugtmkbdfil834 wrote:
         | A cynic in me is amused at the yet unknown corporation being
         | placed under investigation due to a trigger phrase in one of
         | the meetings transcribed incorrectly.
         | 
         | Your point is worth reiterating.
        
           | Terr_ wrote:
           | Or poisoned-data that sets up a trap, so that a system will
           | later confabulate false innocence or guilt when certain
           | topics or targets come up.
        
       | nashashmi wrote:
       | The most depressing realization in all of this is that the vast
       | treasure trove of data that we used to have in the cloud thinking
       | it was not scannable even for criminal activity has now become a
       | vector where we shall have thought police coming down upon us for
       | simple ideas of dissent.
        
         | AlexandrB wrote:
         | A lot of people tried to sound the alarm. It's not "the cloud",
         | it's "other people's computers". And given that other people
         | own these machines, their interests - whether commercial or
         | ideological - will always come first.
        
           | tokioyoyo wrote:
           | To be fair, most people understand that risk. It's just it is
           | very convenient in a lot of scenarios and some businesses
           | might have not even started without it. Privacy is not that
           | big of a concern for a big chunk of people. And they're
           | basically voting with their wallets.
        
       | blueblimp wrote:
       | > Yet this approach is obviously much better than what's being
       | done at companies like OpenAI, where the data is processed by
       | servers that employees (presumably) can log into and access.
       | 
       | No need for presumption here: OpenAI is quite transparent about
       | the fact that they retain data for 30 days and have employees and
       | third-party contractors look at it.
       | 
       | https://platform.openai.com/docs/models/how-we-use-your-data
       | 
       | > To help identify abuse, API data may be retained for up to 30
       | days, after which it will be deleted (unless otherwise required
       | by law).
       | 
       | https://openai.com/enterprise-privacy/
       | 
       | > Our access to API business data stored on our systems is
       | limited to (1) authorized employees that require access for
       | engineering support, investigating potential platform abuse, and
       | legal compliance and (2) specialized third-party contractors who
       | are bound by confidentiality and security obligations, solely to
       | review for abuse and misuse.
        
       | lowbatt wrote:
       | Maybe a little off topic, but is there a way for a distributed
       | app to connect to one of the LLM companies (OpenAI, etc.) without
       | the unencrypted data hitting an in-between proxy server?
       | 
       | An app I'm building uses LLMs to process messages. I don't want
       | the unencrypted message to hit my server - and ideally I wouldn't
       | have the ability to decrypt it. But I can't communicate directly
       | from client -> LLM Service without leaking the API key.
        
         | simonw wrote:
         | "But I can't communicate directly from client -> LLM Service
         | without leaking the API key."
         | 
         | There is a way you can do that right now: the OpenAI WebRTC API
         | introduced the idea of an "ephemeral key":
         | https://platform.openai.com/docs/guides/realtime-webrtc
         | 
         | This provides a way for your server to create a limited-time
         | API key for a user which their browser can then use to talk to
         | OpenAI's API directly without proxying through you.
         | 
         | I love this idea, but I want it for way more than just the
         | WebRTC API, and I'd like it for other API providers too.
         | 
         | My ideal version would be a way to create an ephemeral API key
         | that's only allowed to talk to a specific model with a specific
         | pre-baked system prompt (and maybe tool configuration and
         | suchlike) and that only works for a limited time and has a
         | limited token budget.
        
           | lowbatt wrote:
           | interesting, will check that out. thanks!
        
         | michaelmrose wrote:
         | Will such processing be cheap enough to be done by a box that
         | plugs into a customers router to handle such? Would they buy
         | them? Notably not just for this use case but others
        
         | whyage wrote:
         | Check out https://www.opaque.co/
        
       | ozgune wrote:
       | > Apple even says it will publish its software images (though
       | unfortunately not the source code) so that security researchers
       | can check them over for bugs.
       | 
       | I think Apple recently changed their stance on this. Now, they
       | say that "source code for certain security-critical PCC
       | components are available under a limited-use license." Of course,
       | would have loved it if the whole thing was open source. ;)
       | 
       | https://github.com/apple/security-pcc/
       | 
       | > The goal of this system is to make it hard for both attackers
       | and Apple employees to exfiltrate data from these devices.
       | 
       | I think Apple is claiming more than that. They are saying 1/ they
       | don't keep any user data (data only gets processed during
       | inference), 2/ no privileged runtime access, so their support
       | engineers can't see user data, and 3/ they make binaries and
       | parts of the source code available to security researchers to
       | validate 1/ and 2/.
       | 
       | You can find Apple PCC's five requirements here:
       | https://security.apple.com/documentation/private-cloud-compu...
       | 
       | Note: Not affiliated with Apple. We read through the PCC security
       | guide to see what an equivalent solution would look like in open
       | source. If anyone is interested in this topic, please hit me up
       | at ozgun @ ubicloud . com.
        
       | natch wrote:
       | From Apple's document on Advanced Data Protection:
       | 
       | >With Advanced Data Protection enabled, Apple doesn't have the
       | encryption keys needed to help you recover your end-to-end
       | encrypted data.
       | 
       | Apple doesn't have the keys. Somebody else might. Somebody other
       | than you. Also, I think they meant to say decryption keys,
       | although they're probably just dumbing down terminology for the
       | masses.
       | 
       | >If you ever lose access to your account, you'll need to use one
       | of your account recovery methods
       | 
       | "You'll need to use." Not "there is no way except to use."
       | 
       | >Note: Your account recovery methods are never shared with or
       | known to Apple.
       | 
       | "shared with or known to Apple." Not "shared with or known to
       | anyone else."
       | 
       | The encryption is there, I believe that. I just don't know how
       | many copies of the keys there are. If the only key is with me, it
       | would be super easy for Apple to just say that. I believe that
       | they have said that in the past, but the wording has now changed
       | to this hyper-specific "Apple does not have the key" stuff.
        
       | rglover wrote:
       | > We are about to face many hard questions about these systems,
       | including some difficult questions about whether they will
       | actually be working for us at all.
       | 
       |  _And how_. I 'd lean towards no. Where we're headed feels like
       | XKEYSCORE on steroids. I'd love to take the positive, optimistic
       | bent on this, but when you look at where we've been combined with
       | the behavior of the people in charge of these systems (to be
       | clear, not the researchers or engineers, but c-suite), hope of a
       | neutral, privacy-first future seems limited.
        
         | ActorNightly wrote:
         | Given how politics and companies evolved, I actually trust
         | those people in charge of XKEYSCORE systems more than ever.
         | They may wear suits, but those people usually come from some
         | military background, and have a sense of duty towards defending
         | US, from threats both foreign and domestic, and historically
         | have not really abused their powers no matter what the
         | administration is. XKEYSCORE for example, wasn't really about
         | hacking people, it was just about collecting mass metadata and
         | building profiles, well within the legal system, and the blame
         | should be on the companies that didn't provide privacy tools,
         | because any big government could have build the same system.
         | 
         | Meanwhile, the anti anti-establishment Republican Party since
         | 2016 who cried about big tech turned out to be the biggest pro-
         | establishment fans, giving Elmo an office in a white house and
         | Zucc bending a knee to avoid prosecution.
         | 
         | With these new systems, Id rather have smart people who only
         | work in US defensive forces because of a sense of duty
         | (considering they could get paid much more in the private
         | sector) in charge.
        
           | schmidtleonard wrote:
           | > well within the legal system
           | 
           | It's not a search if we don't find anything, and it's not a
           | seizure if we charge the money with the crime. These are
           | court approved arguments, so they must be correct
           | interpretations.
           | 
           | Point is: modern bureaucrats have proven that they are
           | absolutely willing to abuse power, even in the best of times
           | when there is no real domestic political strife.
        
       | jrm4 wrote:
       | "The goal of encryption is to ensure that only two parties, the
       | receiver and sender, are aware of the contents of your data.
       | 
       | Thus, AI training on your data breaks this, because it's another
       | party.
       | 
       | You now don't have encryption."
       | 
       | Thanks for coming to my blah blah blah
        
         | dialup_sounds wrote:
         | The article has nothing to do with model training.
        
       | peppertree wrote:
       | Is embeddings enough to preserve privacy? If I run the
       | encoder/decoder on device and only communicate with server in
       | embeddings?
        
       | klik99 wrote:
       | > You might even convince yourself that these questions are
       | "privacy preserving," since no human police officer would ever
       | rummage through your papers, and law enforcement would only learn
       | the answer if you were (probably) doing something illegal.
       | 
       | Something I've started to see happen but never mentioned is the
       | effect automated detection has on systems: As detection becomes
       | more automated (previously authored algorithms, now with large AI
       | models), there's less cash available for individual case workers,
       | and more trust at the managerial level on automatic detection.
       | This leads to false positives turning into major frustrations
       | since it's hard to get in touch with a person to resolve the
       | issue. When dealing with businesses it's frustrating, but as
       | these get more used in law enforcement, this could be life
       | ruining.
       | 
       | For instance - I got flagged as illegal reviews on Amazon years
       | ago and spent months trying to make my case to a human. Every
       | year or so I try to raise the issue again to leave reviews, but
       | it gets nowhere. Imagine this happening for a serious criminal
       | issue, with the years long back log on some courts, this could
       | ruin someones life.
       | 
       | More automatic detection can work (and honestly, it's inevitable)
       | but it's got to acknowledge that false positives will happen and
       | allocate enough people to resolve those issues. As it stands
       | right now, these detection systems get built and immediately
       | human case workers get laid off, there's this assumption that
       | detection systems REPLACE humans, but it should be that they
       | augment and focus human case workers so you can do more with less
       | - the human aspect needs to be included in the budgeting.
       | 
       | But the incentives aren't there, and the people making the
       | decisions aren't the ones working the actual cases so they aren't
       | confronted with the problem. For them, the question is why save
       | $1m when you could save $2m? With large AI models making it
       | easier and more effective to build automated detection I expect
       | this problem to get significantly worse over the next years.
        
       ___________________________________________________________________
       (page generated 2025-01-17 23:00 UTC)