CTRL/SHIFT - 13 06 2026 PARTHENOPE STAGE - GIOVANNI MAZZEO - FROM MISINFORMATION TO VERIFIABLE TRUTH

I want to introduce the next speaker Javanni Mateo, associate professor of computer science at the University of Naples Partanopa and he is going to talk today about AI generated content which is becoming indistinguishable from reality and here comes Atestia a decentralized protocol designed to restore the trust in digital media through an open verification network. So Javanni the stage is yours and please. Thank you.

Okay. Thanks. Perfect. Okay. Thanks.

Thanks. Thanks a lot. So uh yeah, my name is Joavanni Mateo. I am associate professor uh in computer security at University of Naples Partnen and uh I am co-founder of this company called Atestia. I am specialized in the field of cyber security and blockchain.

I had different consulting activities in companies in the web3 space and uh and in particular in in fintech and yeah so as you have seen from the title uh today I'm going to talk about the problem that we are used to face uh regarding manipulated media which has a an important impact on our society. We have a situation where there is a trust problem in our internet. So some time ago we had much more trust on what we were seeing on on our internet. Of course there was the risk that uh someone could manipulate the media but in that case was just uh possible from someone really specialized which could use for example Photoshop in the right way as a pro and uh and so it wasn't that easy to produce uh manipulated content that could be images, video, audio. uh while today with as you know with the generative AI it's much much easier with the recent deep fake algorithms it's much much easier to create manipulated media deep deep fake content that in certain cases looks very real.

It's true that sometimes we are used to see these pictures on social media wherever on our news and in some cases it's really evident that these are fakes but it's not always the the case and also the trend is uh not actually very good because it looks like the the quality of the content that is being generated is improving a lot years over years. But the fake content if you think about it the fake content is only half of the problem because what we have as a consequence is that also the real content can be becomes deniable. So it's uh a problem that we cannot know at a certain point we could also ask ourselves about real content and if that real content is actually uh real right so there is the need of something uh of an infrastructural solution for this kind of problems because it's getting worse and in fact the deep faith detection uh market is increasing. companies are aware that is not anymore a nice to have a feature. But those companies that usually uh manage content and and media, they have to they need this these kind of detectors that as a consequence there there are more and more billions that are circulating in this field.

Now you could ask okay there are already solutions for defect detection. Not sure if you some some of you already uh used one of these solutions if you have experience in testing these tools and it's very frequent that these tools provide a a wrong result. So you send them a an a real picture they give you an accuracy of 60 70% that that real picture is AI made there are it is very frequent that these tools fail why is that first of all there is so why these tools are not enough first of all these tools are opaque so we don't know we have no uh actual knowledge about the models the trained data uh they are pretty closed. Then it's also uh one problem is also that they these tools are trained on uh data on different types of data but it's uh pretty uh it's well known that the actual the the quality of the detection depends a lot on the same on the on the data set that was used. So I mean if I have to detect an a an image containing animals and uh I have to check the quality of that image.

So my model has to be trained on this that kind of data right and we it's not that easy to have a a learning path on on all these data set. Uh so this these tools are fragile. uh as I as I've said as I've se as I said before uh they fail very frequently. So there is this missing layer. We have yeah identity that can tell us who created or shared something.

We have also solutions for provenence that are not the panacea. So I mean provenence can tell us where content came from but it cannot tell us it is attackable is not perfect. So there is even there a problem. Then we have tools for detection but we still need a reliable uh auditable uh proof of authenticity. So what is Aestia?

A testia is a solution that is uh trying to provide a decentralized infrastructure for detecting manipulated media. And so as you can understand is using is based on on a blockchain and is so is decentralized is verifiable because all the proof are actually stored on chain and we use zero knowledge proof to give insurance on the aggregation of verifications is also incentivized. So the idea is that we want to create a economic incentives to give stability to a network of verifiers that should participate in the evaluation of content. So we receive a score, we uh we receive different scores, we aggregate them and based on the quality of the uh of the consensus, we could either uh reward or slash the attesters, the verifiers. What is the flow?

Just to give you an idea, the flow is that is the following. The the submitter of the content, the contributor uploads a media and at that point an onchain attestation is generated and opening up the verification window. Then our detector provides our own evaluation and different independent evaluators part that are part of the network provide their own score. And these scores are all uh provided as offchain attestation signed by the different uh attesters the different verifiers. The scores that are then aggregated and combined and this process of com of com of aggregation is proven via zero knowledge proof and finally the aggregated score is anchored on chain.

So the the final outcome is uh I mean is auditable and is trackable. At that point based on the on the stage of the network based on the situation the the attesters are rewarded or slashed based on their performance and also on their reputation. Of course there is there are issues. So it's not that easy to uh it's not straightforward to set this up. There are problems that maybe some of you already thought thought about.

First of all, there are code start problem. So uh the the quality and the uh the trust in the network strongly depends on the amount of attesters and verifiers that participate to the network. the the network quality. So of course the network should attract qualified verifiers and there are risks related to the consensus. So if few testers participate the aggregate score of course may be fragile.

And then last two bullets, the risks from coordinated groups that could manipulate, could sync. And even here we have a situation where the higher is the number of attesters and verifiers, the better is the quality of the of the network. And last but not least, there is uh also it's important that the network could reward and slash on the actual performance and not on content that is not uh uh that was actually real or made by AI. So in terms of use cases uh where these kind of solutions are very important these are very important uh for example if you think about in legal context so legal offices that require these a a proof of on the quality of this content also on from in the con in the space of media and journalism we need because I mean before publishing a specific article a spec a specific image it's important that we uh that the the uh that the journalist is sure about what is actually publishing and other example are for example from uh insurance right in the or in the also in the financial space in insurance space someone could use a fake media to to claim uh false statements. And so this is what where we see the actual importance of our solution.

And with this slide ah okay I think this last slide is missing. Okay with these slides I I'm I'm done. I think one slide was okay. I think the last slide is missing. And uh yeah, with this slide I closed uh my presentation.

Thanks a lot for for the attention.