# Immutable Future: Crafting Trustworthy AI with Blockchain - Kosa Nenadic | Ethernal

- Channel: [ETH Belgrade Community](https://streameth.org/eth-belgrade-community)
- Date: 2025-10-07
- Duration: 19:46
- Topics: People & Blogs
- Watch: https://streameth.org/watch/yt-4Jihxq8xSLs
- YouTube: https://www.youtube.com/watch?v=4Jihxq8xSLs

## Description

Immutable Future: Crafting Trustworthy AI with Blockchain - Kosa Nenadic | Ethernal

## Transcript

Hello everybody. My name is Kosa from internal and today I'm going to talk about immutable future. How to craft trustworthy AI with the blockchain. So let's start. In this presentation I will talk about AI. I will explain few definitions and AI concepts. I will go through core blockchain features. Then I will join these two worlds together and explain how we can benefit from blockchain in AI. Afterwards I will say few words about our enterprise blockchain that we have built at Eternal and it and give few use cases how we can apply uh blockchain in different AI processes. In the end I will give a conclusion about this topic. So let's start what artificial intelligence is. It's the capability of computational system to perform human intelligence tasks. These tasks include learning, reasoning, decision- making, perception, language understanding, problem solving. And at the same time, artificial intelligence is a field of research in computer science that develops and uh studies methods and software that enable machines to act intelligently. Today we are living in a world where popular large language models are overwhelmed. let's say they're used across different products and uh with them um AI systems are becoming increasingly complex and autonomous and at the same time these uh AI systems are adopted across industries and they are affecting decision making processes and that is why we were we have to be very careful how we use the AI especially if we re we rely on results of the AI AI. Uh the other thing is machine learning. Machine learning is a sub field of AI. It basically give us um ability gives computers ability to learn without being explicitly programmed. Here on this diagram we can see what machine learning does. We have a machine learning program or system that has training data which is input data for training. We have a basically plain model and after training is done we get the training model with u parameters that suit to training data. Once we have a trained model we can leverage that train model to derive results predictions to make predictions about new data based on the train data. So this is important. We have a new data. We have a train mall and we have results which are predictions. Now we come to generative AI. This is buzzword these days. And what generative AI is? It is basically AI that creates new original content in response to prompts. These prompts here are basically new data that we saw previously on diagram. Here in this image we can see relation between different parts of artificial intelligence. We have a machine learning which is sub field of artificial intelligence. We have a deep learning which which is subset of machine learning. And here we have a generative AI which is basically subset of a deep learning and a large language models that are also part of the deep learning but uh they are not the same. there is a difference between large language models and generative AI and that difference is basically um with the LLMs in the multimodality because well-known models like uh GPT gemini code are large language models that can serve as not that can serve well which are empowered with the generative AI so they're in intersection of these two This multimodel large large language models can understand, process and generate multiple data types like text, imagery, audio, video, video, code, structure data, etc. And now that we uh have these models here which basically create original content which is new, we have to be sure that the outputs of AI are trustworthy and there there is a catch essentially. So what happens in uh in AI models that is that we have a number of anomalies. Here are several categories of anomalies that we have. They include data anomalies, model behavior anomalies, training, inference, security, ethical life cycle anomalies. And as we can see these each category has a number of different types of anomalies. I will just mention a few uh for example biased data biased data sets can lead to biased results. for example that it will um bring like systemata systemic prejudice in the results and once we have something like that that was that would also uh affect ethical and social aspect of the result because we would introduce uh stereotyping. Another anomaly is for example hallucinations. This is something that is uh frequently mentioned these days because uh AI models generate content and although that content looks um reasonable, it is um factually incorrect and you have to have a wide knowledge uh critical thinking uh every time uh when you analyze the results or when you want to rely on the results of uh for example large language models. I can also mention prompt sensitivity. This means that uh when you change inputs for for the model uh you get uh very different results for for the small changes in the input. Um okay we can continue now that we have saw what um AI is and what are its parts. Uh let's see uh what is a blockchain. So we are you are all already familiar with the uh blockchain. Uh it is a distributed uh ledger. Uh basically it is a long list of cryptographically linked uh blocks. Uh core features are the features that are important for this let's say merge between uh AI and and the blockchain. So core features are decentralization, immutability, transparency and cryptographic security. Decentralization is basically distribution of control and authority across all network participants. Immutability means that once uh uh once once transaction is recorded in into a block on the blockchain, it can be it cannot be uh deleted or altered by any single party. uh transparency. All transactions are visible to to participants who have access to the network and cryptographic security. Data on a blockchain is secured through cryptographic techniques uh which make uh which makes u it um which makes um unauthorized access un u unavailable. uh and it's also enables uh uh enables that the blockchain is resistant to tampering. Uh one thing that also has to be mentioned is that um we have um programmable infrastructure on a blockchain for example on EVM blockchains we have u um smart contracts on other blockchains we have programs etc. And this is I mentioned this because uh in examples uh that uh that we that we will see uh we will use smart contracts. Um now that we have sold these two worlds um we basically have a two sophisticated technologies which operate on different paradigms. The first one AI involves adaptive learning processes and it includes probabilistic logic while on the other side we have deterministic and immutable uh ledger. So these both technologies uh are reshaping the industry today and uh it is interesting to see how they can be let's say joined uh and what opportunities it can make. The key challenges of AI include u transparency, accountability and data integrity. And we will see how blockchain can solve these issues with the core features. Now that we recaped about blockchain, let's mention our blockchain. This is a blade. It is a basically cutting edge private blockchain platform uh which is design designed with a clear focus on enterprise needs. It is basically EVM compatible blockchain which supports typical uh Ethereum blockchain features. Uh on the top of that we also offer um enhanced features like uh improved scalability, faster transaction processing, free transactions when it is require required and the late the thing that uh we have de developed recently is a built-in bridge to EVM compatible networks. And now let's go to the use cases. I will present several use cases which show how we can use blockchain in AI processes. The first one relates to data set preparation. The second one uh is related to enhanced AI model training. The third one is secure and verifiable AI inference and the fourth one basically incomp um joins previous together. So in this figure we can see how data set preparation process uh looks like and what uh is added on the top of it to uh integrate blockchain. Uh on the on at the beginning we have u inputs. These inputs are usually taken from the internet. Uh they are taken from the web as web scrapes. Then we have a user data sets or for example uh data that is taken from IoT devices. Once that we have that data that data comes in uh different formats and what we have to do is to pre-process these uh inputs. Uh as a result we get data set metadata and we get a data set. uh this data set is basically pre-processed where we we have done uh pre-processing uh like filtering, normalizing, augmenting and when we have these two we can create a hash of data set and a hash of met metadata and store those hashes on a smart contract on on a blockchain. Here for example we have a data provenence smart contract where these two hashes are kept metadata uh data set uh takes care of information about the sources for the data set um what type of processing was done um who is the owner of the original data who is the owner of data set these the data set and the metadata are kept usually offchain. Um okay. So what can we uh we can also say that uh this data sets can be large. So in order to create hash we have to do something like splitting data set in smaller parts and creating hash of that those parts and then applying some cryptographic technique like creating a Merkel try of the hashes to create a Merkel root um hash and store that hash here on on the blockchain. Okay, now we have a data set. After that data set is created, we can leverage data set and other data sets to train a model. So the process start with the uh regeneration of a data set hash. Why we do that? We want to ensure authenticity of the data set. That is why we take hash newly created hash and check it on the hash that is stored on blockchain. After the verification was successful then we have then we can continue with the training of of of our model. When model is trained, we can create a hash of that uh trained model and at that time we can register a mo model hash on the smart contract the other smart contract on the blockchain and at the same time we can we can register an event describing the process of training. Uh basically this event includes model version, owner of the data set, owner of a trained model, for example, access rights, uh what data set was was used for training or data sets were used for training and similar stuff. Okay. Now we have a verify data set. We have a we have a train model that is uh stored on the blockchain. Now we can um infer new knowledge from our trained model. Um now the first step is to do verification of the hash of a trained model that we want to use and to give input to our train model. This is inference data. These this is like a new data from the first one of the first slides. Then we go into inference process and we get output of that process which are which is some result. When result is created we generate a hash of the result and and the proof. Both hash and the proof are also stored on a smart contract on a blockchain. Why we do that? We want to be uh we want to take care of um uh data integrity of the result. At the same time we want to prove that how training was done uh not how inference was done from the inference data and what was what was the result. At the same time this proof uh can hide the privacy uh private data that that were used in this inference process. So we can for example use ZK proofs here and store everything on on a AI inference uh smart contract. Okay. Now uh this is the last uh diagram which basically shows DI application inference process where basically we encapsulate um everything that we uh saw. We have a user which uh enters decentralized AI application. Uh it authenticates through wallet. Basically through this application user selects a model that he wants to use. Uh creates prompts that he wants to use with the train model. Basically call calls inference one. Once this transaction which calls inference is called with its inputs uh transaction log records that uh transaction request and decentralized compute node read events. Once the event is recognized by this by this compute node, it fetches appropriate model. That model is verified on on the blockchain and it is executed with the inputs that were given by user. Results are saved uh in the result storage and this everything happens on on offchain. uh when result is saved at the same time we create a result hash and a proof and these are stored on the blockchain and when this is done user can uh use this result hash to retrieve results from the storage. So with this approach basically different phases in AI uh in different phases of AI processing let's say uh are captured on the blockchain. We we saw data set we we saw a training model we saw inference and now we we also see uh user calls. Okay, now that we have gone through all this, we saw that AI has a complex challenges uh which include transparency, accountability uh and data integrity. Uh and we also uh saw what the core features blockchain offers. With these um core features uh we basically address uh AI concerns and um opportunities that uh were created through this join um are for example enhanced data security and integrity. We also have a improved model traceability and version control. We have a transparent and verifiable audit rails. We have execution on a smart contract and we have a reliable provenence and compliance. So we can uh suit this processes to fit into different regulatory rules. So we can comply with everything that is needed. And one thing that I also have to mention is that we have recently created blockchain in AI use cases white paper. So you can find more details about this topic in this white paper. We have developed it uh together with the cybergate defense recently and you can check it and that would be all from me.