Signal vs. Noise: How LLMs Broke the Internet and How ZK Proofs Are Going to Fix It |Tracy Livengood

you connected good awesome thank you everybody uh my name is Greg over the MC for the next last section here and today we've got Tracy who's going who's from Pluto a ZK Dev tooling team and he's going to talk about a very interesting talk today about signal versus noise and I'll let him take it from here all right hello hello can you hear me all right a little bit closer all right that looks good hi everyone um so I'm Tracy like you said I'm the founder of a company called Pluto we're building a Dev platform for applications to add cryptography uh which we think brings novel features to those applications um today I'm going to talk about this idea that's been on my mind for a few months now which I call signal versus noise and really the problem is um in the last 5 years we've seen a rise of llms and generative AI that can produce content that's more or less indistinguishable from Human content um and I'm going to talk about the solutions that exist today and possibly some ways that cryptography can help with that and this really started as an exploration into to Identity and fraud more than anything but the deeper I dug the more I thought about misinformation and how important that is to solve as well okay so first off I'm going to go through a few examples that are kind of coming up recently so first is real ID real ID fake ID today's kyc and identity Solutions like stripe jumio Persona they're built on the notion of a valid identity um and generative models have reached a point of quality that they can construct those IDs that are really difficult to uh tell from a real identity so this is actually an llm or like a generative AI model uh identity um and it was used to bypass some okx uh identity um kyc so there there are still safeguards so for example this has on it like an issuance ID and you can actually check with a central Authority like Lexus Nexus to ensure that the ID was issued by a government organization but um um if you mix this fake ID with like a valid serial ID you can actually bypass those checks as well so this is a a challenge next is real person fake person so text generated from an llm at this point is just indistinguishable from Human text and that's also becoming true of voice and video and the cost and sophistication uh to create these fake people is is decreasing and then finally real media fake media so like I was saying generative AI produces increasingly realistic video audio images and we can no longer trust our human sense to help us determine the truth um so imagine this scenario you're watching a video on YouTube and you think to yourself damn that can't be real like how would Donald Trump say that uh then you attempt to investigate you Google like did Donald Trump say this uh ultimately despite your best efforts you literally cannot tell whether it was real or not and you give up and you kind of move on with this weird sense like I don't know the truth here anymore I'm curious how many of you have had this experience already yeah that's nearly everybody so just broad Strokes like what is the um the landscape there's three big problems it's fraud spam and misinformation fraud is typically financially related so that's about kyc and KYB and those are the techniques that we use today and the regulations associated with them um spam is about creating fake accounts and using them to misuse a platform and misinformation is about spreading information that is dis difficult to distinguish from real information okay so what Solutions do we have for fraud today like I said earlier we have stripe identity Persona plaid Lexus Nexus and they're very sophisticated platforms it's not that they haven't given it a good go um each of these platforms employs many many techniques so uh today they'll use things like deep neural networks uh based on videos of a user holding up their their license and the actual goal of that interaction is not just to capture the license text they're actually monitoring to see the reflection and the sheen on the license the person picking it up and putting it in front of the camera all of that information is collected and it's part of determining that it's a real ID um they also use biometric data so often if they're asking for a webcam they're actually looking at your eyeballs as well to try and see if you're a real person you're looking around the room um they monitor how your mouse moves around a website um they check that pared identity against a central database um they use really adaptive ml techniques so emerging threats are kind of constantly happening in this space it's a game of Cat and Mouse um and when a new technique comes up often it's so hard to respond for these companies they'll just like write a hard rule like okay we see a new technique you know it's coming out of Slovakia it's like breaking through our our checks here like let's just ban everybody in Slovakia for the next like week and until we can figure out a better way to to fix this problem and they build those into like a programmable decision-making framework um what about spam and misinformation so a few Solutions exist for spam and we've seen this on Twitter that's like account gating paying $8 to get a single account or civil resistance mechanisms like capture and also content moderation um like the infamous horrible job that a lot of Facebook uh people have sitting around and looking at terrible content and then on misinformation there are things like Community notes so having the community try to digest information and determine if it's real Google fact check and a coalition for Content Providence which is an attempt by like content creation tools to create a chain of trust from where you created a video to uh your eventual distribution of it cool and then briefly the regulation landscape so there are Global regulations in every country that try to uh put a framework in place for these problems um here's a list of them I'm not going to go too deep on regulation but it'll be in the slides if anyone's curious to look okay so really I'm posing the question like is there a problem is this problem solved or is it an open problem still um and my case is essentially like fraud prevention is pretty robust today um but despite that the rate of fraud continues to increase so platforms already have a really strong economic incentive to get this right there's multiple billion dollar companies is trying to solve it um and the costs are sort of directly measurable when the fraud rate increases the revenue decreases um so that one I think we're doing an okay job but it is getting more challenging then on spam and misinformation I think this is where there's really an opportunity to do a better job so historically misinformation is focused on text as I said generative AI makes highly sophisticated audio and video misinformation possible and today's Solutions are really built platform pipe platform only by the largest organizations in in traditional internet Google Facebook Etc um they essentially have teams viewing this content trying to train models to predict whether it's good or bad information and information is highly subjective so it's very difficult to determine a sense of ground truth furthermore because they're so specific to a given platform like like say Facebook um they're highly centralized and it's very difficult if you're a small blog to use this type of solution and have the similar quality of content moderation as as Facebook might and so I propose a few things that are missing from what's out there today um first is anonymity preserving tools for fraud so today's fraud requires you to reveal fraud prevention requires you to reveal who you are and um that involves requiring your revealing your pii data to some Central Authority um I think we're probably going to enter an error of the internet where almost everybody who uses it has to be kyc to avoid the spam that's going to to arise from from AI the downside of that is it's a very dystopian world to imagine that all internet activity is kyc um so I think we need Anonymous preserving tools that allow you to use your identity to prove you're a human but also give you privacy and anonymity I also think we need Universal tools and platforms for misinformation so as I said there's no universal truth uh on information it's subjective so each platform has to make judgment decisions but I think every platform needs access to a tool for making those decisions and then I also think we need a general solution for spam but it's almost impossible um spam and not spam is is another subjective uh challenge spam's only detectable when it becomes abuse the first message from a user might not be spam but the second might okay so what what are some emerging techniques like this is all valid pain points but we haven't really talked about any solutions so there's for those of you who are in crypto there's uh quite a few that are up and coming so one is proof of humanity um this is World coins sort of scan your eyeballs um determine that you're a real human give you an identity and then let you link it into a platform um there are challenges here there's black markets for identities people in third world countries may not mind selling their like retina scan for three or five bucks um and there's also privacy concerns excuse me we're trusting worldcoin a lot to use that information and protect it um and despite their best efforts you may have a security leak and in the event of some exploit and their stack like extremely sensitive personal information can end up on the internet and never be put back in the box so I think it's a legitimate concern um I don't know if those of you have paid attention to uh uh the DNA processing 23 and me they recently had a scandal where they were selling like DNA sequences to third parties which is kind of like permanently revealing information about you to the to the state or whoever's buying that information uh next possible techniques and solutions are signed identity documents so I don't know if everyone knows this but inside your passport is an NFC chip and on your NFC chip is assigned attestation from the government uh private key about your information so this can actually be a really powerful source of identity verification you could tap your passwort on your phone extract that signed information and have a pretty robust um uh source of identity and then the final is just composable attestations and reputation so if you can take these attestations from a few sources proof of humanity a tapped passport maybe some uh personal information about where you live your tax returns Etc you can compose these things together and maybe layer on some social reputation you might be able to make a pretty robust artifact about uh your individuality and your proof of humanity and then on misinformation some emerging techniques so one is ATT tested sensors that's microphones cameras location data um you can actually in the hardware of the device create assigned uh production of content and use that signature to verify that it came from a trusted public key so influential figures are probably going to start publishing most of their content from signed devices only um and recently we saw the White House announce um they'll start doing this in 2024 um um that helps but it doesn't actually solve the whole problem a big part of content production is the initial production but then the editing and modification of that content so we've also seen a few cool hacks around verifiable modification of content cropping resizing images um while maintaining proof of Providence so checking inside of a ZK snark that it came from a signed Source transforming it inside of a snark and then producing a proof that it was transformed correctly and it was the signed data that was input um there was a cool hackathon project at Stanford hackathon for this um and another one is this cool idea called perceptual hashing which is taking an audio or not necessarily audio a video or photo and hashing it to check the similarity to a prior video or photo before cropping screenshotting or modification um there's a paper from Apple called neural hash um and it offers relatively robust hashing across Transformations so you might take an initial image screenshot a piece take a it out and then compare the prior and the latter perceptual hash and get like a 90 plus% positive ratio um which can help to detect that the content derived from a a valid source and there's actually a team Morpheus Labs who just did some work on this a few days ago cool how do these all help so really if you think about it like I said at the beginning the goal is to only have 8 billion human actors on the internet at least in a setting where other people are consuming that content it would be great if you could sort of binary know it's coming from a human that's a valid source and if you could from these actors derive a fair reputation score um you may be able to remove some of the noise from the internet um it's also important like I was saying that the classif classification is robust and forgiving if we have one person online and somehow they get misclassified it it's like taking away a fundamental human right so there needs to be some path for remediation uh from that situation and then from this set of actors you can sort of classify strong versus weak information so I took a photo from my attested sensor I have my identity linked to this ATT tested sensor and I can prove it strong information from me um and weaker information like deriving perceptual hashes to show that I had the content I modified it maybe put it in a meme and then put it on a website um and the ultimate Challenge really is to do this while preserving anonymity privacy and freedom if we build these systems and they don't have hard cryptography at the four of them uh it ends up being a pretty dystopian internet that I don't think any of us want to use so individual Freedom needs to be protected for this to actually work and then maybe I'll spare one one thought for the AIS which is I think that AI generated content in general will be an important part of the internet um I'm not saying we should ban all AI generated content it's just a question of being able to differentiate the two and know the Providence of some content and then finally can blockchains help with this um I think the answer is yes I think blockchain serve as a great shared protocol source so as I talked about before Facebook has their own platform for doing this problem but it's very hard if you're an individual content creator to use that platform if we could have a shared protocol for identity reputation content Providence and we could use that as a source of Truth um I think it would be a better tool in general than one centralized system for this cool so the path forward is three problems the Civil problem build a chain of trust from content creation to distribution and create shared protocols for information and user quality and that's that um how we're thinking at Pluto um today we're building tools for the internet and developers uh we want to keep the internet weird but keep it useful and we're building trustworthy techniques for merging the internet and blockchains and using Advanced cryptography to create a transferable chain of trust all right questions I see one hand up do we have microphones first I want to say I really love this I mean I think this is super important to stuff to think about um as far as like the composable stuff I I love that too have you thought about you know everybody it's you know we're in the decentralized world right so it's like one guy can't be the center of trust for the whole world have you thought a little bit about how to interoperate between different like centers of trust uh yeah I have I think a big part of what I describe as composability is using programmable cryptography to verify an attestation from some arbitrary source and kind of like boole andand it together with attestations from other sources in the beginning you could imagine a heris model so you have like if conditions if this then that if this then that um but you could also get to a richer model um where maybe you're doing a binary classification with a ml based estimator and you can also verify that inside of a snark so oh there's one more testing there we go H there hello thanks for a great talk super interesting um I was wondering if you talked about almost like putting a watermark on content you speak a little louder I can't quite hear you yeah is that better yeah okay you were talking about putting almost like a watermark on some content if it comes from the White House or something like that so that people can see that that came from them yeah but there's no watermark on any AI generated content so there's no proof of the Providence for that yeah would another alternative be somehow regulating the AI generated content to have Providence on chain so that you can see where that content originated from in the first place I see um so one I'm not a I'm not a fan of trying to regulate things um I think it's even if you tried to make it the case that you had Providence for any AI generated content um from like one fixed Source someone could just spin up an open source model and start generating content so it's very difficult to have guaranteed Providence of content um I think the best we can do is uh have signatures and attestations to what we know is good valid content from good valid sources and the rest we essentially have to consider noise um and then you have to opt into how much of that noise you want to experience so um for like White House content yeah I think it's about signing the content and any user who ultimately consumes that content being able to look at the signature verify it themselves and see that it actually came from the White House having like a little check mark next to it um um and then if users want to edit that content maybe like remix it into a news feed or something I think it's about having segments of that video that can be verified as derived from an original attestation and that's kind of where perceptual hashes I think come into this hello um what incentive do you see for uh centralized parties and Commercial parties to um use decentralized solutions for uh provenance and other things is it legal is it commercial something else see yeah the so I don't think there's an incentive for a central organization who's already doing kyc or identity checks I think they'll continue to use centralized technology um but I could imagine a world where platforms that value anonymity uh find these Solutions the most appealing to avoid having to kyc and manage that type of identity information so I think Twitter or Reddit where like famously anonomous profiles are a key part of their social strategy um would like to find decentralized solution to this that's that's my hunch okay that's time thank you everyone for coming I'll be around for a few minutes if you want to ask questions