 So, my good morning and welcome to this open air nexus webinar. Before we, we go on with the webinar I just want to go through some housekeeping rules that you are already used to see so this event will be recorded as you, as you will see the participants be asked the participants to have your microphones off. If you want to participate user shot to introduce yourself to interact with the participants or dress questions to the speakers. At the end, you can always raise your hand to speak open your microphone and your camera and address your comments or doubts to my knowledge that will be here to to help you. The presentation and recording there will be shared with you by mail but we also uploaded in Zenodo and YouTube, and we'll also update it in the open air portal, and do share your this webinar through our social media on tweet. You can use our hashtag open air you or open air underscore amnesia. So, today, we have my knowledge, there are, there are vitis from my Tina research center that will present the amnesia tool I accuracy data anonymization tool that helps you to transform personal data to anonymous data. And, and now without further ado, I'll just at the end of the webinar sorry for that at the end of the webinar we will have a Q&A session where Manolis will try to to address your comments and I'll send suggestions. Okay. And now I'll pass the floor to Manolis. Thank you so much for accepting invitation to be with us here. And the floor is yours. And thanks a lot Paula for inviting me and for the introduction. I'm happy to be here and presenting amnesia which is a work we have a project we have worked for a long time. Maybe some of you have already seen it died or have heard about anonymization but I will try the presentation to start from the beginning and explain when what data anonymization is. So, what I'm presenting is amnesia which is an open source and free to use tool that anonymizes data, and by anonymization women here. In the sense that GDPR does. So, GDPR clears, regulates efficiently. The users of personal data. So, it makes the field clear but also imposes some serious limitations on how to use personal data so you want to use data for personal data for a specific purpose you either have to have a law that supports it or a contract or user consent. It also has a window for is for using them in research but that does not cover any case or any purpose in research or exchanging or exchanging data with other organizations. So, when usually data used for research to be shared for example in a project, they're used based on consent and consent might be difficult to manage and track it might be revoked. It limits your purpose if you find something new, and you might get it to use it but it doesn't remove the dangers of leakage or for the danger of data being misused and the data controller remains responsible for that. What we're trying to do with anonymization and with amnesia is to unlock the potential of data to take the useful information that lies in the personal data outside the outside the GDPR scope. So, a key observation here is that when you usually use data for research in many many cases, if you do not need to actually identify the person, right, maybe when you do medical research and you do clinical study of course you need to know the person if you want to train a model on data from personal habits or in several fields or get statistical information that you only need the statistical information that lies in the personal data and not the identifying information. So, what anonymization does is this transformation from personal data to statistical data by removing the identifying information and preserving as much as possible the statistical information. In this process there is a reduction in data quality, which when you have a small initial data set can be important, the larger the data set grows, then this reduction in data quality becomes less and less important. So, anonymization unlocks the potential of personal data, it removes the constraints that GDPR imposes on them. I need to stress here that anonymized data will be always be different from the original data. So, anonymization is one way transformation of the data. What we provide here is actually a statistical guarantee that no one can reverse this transformation. So, this makes it GDPR compliant and guarantees that the resulting data will not be identifying. And therefore also to highlight here that pseudonymization and anonymization are different things. In everyday language, in everyday practice, we use the term anonymization to cover usually both cases. So, anonymization is the removal of direct identifiers and there is the replacement with pseudonym or even no replacement at all, right? Anything can actually act as a pseudonym. The key difference is that when you just remove the direct identifiers, you reduce the risk that someone will be identified, but you get no guarantee that people cannot be reidentified. Okay, I will give you an example for this a bit later. The idea is that if you remove the direct identifiers and you do not pay attention to the rest of the data, then other secondary parts of descriptive information can be used to reidentify a person. So, pseudonymous data and pseudonymization reduces the risk of reidentification but according to GDPR this data remain personal data since you have no guarantee that someone is not reidentifiable. And they lie inside the scope of GDPR with all the limitations of GDPR. This is not the case for anonymization. Anonymized data outside the GDPR. And secondly, we guarantee this is that there is a statistical guarantee about the risk of information leakage. And according to my view, this is the most productive way to share data with third parties that you either don't fully trust or you have not gotten consent in the beginning for sharing them for research purposes. I told you mostly the good things about anonymization but we need to be clear on its limitations. You always lose some information because you need to remove identifying information and the data quality is reduced. There are some great boundaries between anonymized and pseudonymized data. And here is that every privacy guarantee reduces the information risk but it is information leakage but it's not completely removed. So you need to make a study, the DPO needs to decide on the correct parameters and the correct guarantee. And there is room for decisions and statistical guarantees only are models for a social notion called privacy, they're not by definition privacy it's guaranteed covers it in a different way. And at least at this stage it takes some effort from the user's part to anonymize the data because it is a novel process we do not have yet enough information from user experience to be able to offer automations. So for Amnesia a key objective in our development is to make it effortless for the user to anonymize data and I think compared to other solutions that exist were one of the most user friendly. To elaborate more in the scenarios that anonymization works best. I think in the general in if you want to see the different scenarios they're all for inside data sharing you anonymize the data to share them so if you're a practitioner that collects data and then wants to outsource them. So in such purposes anonymization can help you a lot because it reduces the actual risk and it enables you in cases to circumvent a consent or a detailed consent right you might just get a consent to anonymize and then you can do with the result whatever you want because it's statistical data. It's a way to give data to recipients that you know do not fully trust. This does not mean it's you give to someone you don't want to, but you can give it openly on the internet for everyone to write your not obliged to get an NDA or a contract with them. And this really simplifies with technical means a process that might be quite difficult from an administrative point of view. I also like to highlight here highlight here the difference encryption encryption methods protect the channel. They protect the data transfer, but the recipient of encrypted data usually gets the key to decrypt them and process them so encryption can help with the security and confidentiality of data sharing, but it cannot help you. From the data recipient right you need to trust the data recipient. And the last thing is you use an organization when some reduction to date information quality is acceptable. The recipient wants to do some kind of statistical analysis. If not, if no reduction in the data accuracy is acceptable then an organization probably is not the way to go. So why should someone use amnesia. First of all, they have very very few tools that do data and my station in the GDPR sense. There's another academic tool with more or less similar functionality with amnesia. As you have seen in SAP they have some functions for differential privacy, but there's no tool at the moment that's commercial or widely used. All these tools and a bit of explorative anonymization has been used in very few public cases. Okay, the most famous one is the use of anonymization techniques for the publishing of US census data in 2020. But we have seen and with amnesia many companies that are interested for using it, the confirmed cases of any kind of anonymization are still very few. So we with amnesia we try to make anonymization simpler for everyone. So we gave focus on user friendliness and less on having many different guarantees and many different features. So we try to be user friendly and always to improve on the interface. It's private by design, it's not an online service, we have an online service for demo and training purposes but if you want to anonymize sensitive data, you download it to your local computer or local internet and then it works locally so no data leave your premises. And we give several tools to the user to customize the anonymization process. Not many different privacy guarantees but several different options in guiding the anonymization to reduce the information loss. So we are trying to support many different types of data, which is quite important in practice. We're, I think the only solution that supports sparsely dimensional data, this, what we call in databases set value to data is maybe an example is a retail store retail store bill where you have an arbitrary number of transactions from a very large domain. And finally, we have a clear architectural separation between the interface and the anonymization engine. So amnesia anonymization engine is easily integrated, it's easy to integrate in different IT systems and to get it in your own workflow and make the anonymization work seamlessly for the end user. Okay, it's some years now already up, we have more than 100,000 visitors, 400,000 page views and more than 7k downloads. So we have gotten a lot of feedback, bugs have been reduced and it's kind of mature. I don't think how many details on the status need to be given here, we support now canonimity and came anonymity and where we have been a non-public yet version, a different cell private anonymization algorithm, which will be released in the next few months. And we offer also a rest API that allows this integration with other workflows and data management platforms. So it's not that organization does not need to happen through amnesia web graphical interface. And after a few years of use bugs have been diminished and scenarios work without anonymization scenarios work out without problems. Now, I will go back to anonymization itself and explain the difference with pseudonymization. And I will get to a real example that comes from one of the first research papers in data anonymization. And it shows the dangers of pseudonymizing and not paying attention in secondary and descriptive information. Now, on the left side here you can see data that were published in the UK as, you know, statistical data from hospitals for its hospital visits, the hospitals in Massachusetts provided this information. And they had removed the name, the social security number, and they were just saying that someone was born on this date from this, staying at this zip code, and they reported also the section of the patient and then some other descriptive information came and had these diagnosis. Now they considered this safe because there was no name, and there were no direct identifiers, you could treat this data and considered as pseudonymous so they probably had, you know, some number also that identified the record which would be the pseudonym. At the same time, from the voters list, it was available, the information on the right cycle was available, which was the name of a voter, their address, and again they had the zip code, birth date and sex. So, this data was not sensitive and it was public, but a name was there associated with this descriptive information that we call was identified at the same time there was a second database where sensitive information was published without names but with this was identifiers. And by linking this, these two public databases, you could re-identify the patients that visited the hospital so there was a privacy breach that concerned almost everyone who's on the voters list visited the hospital. And these are the dangers of pseudonymization that come through the processing of secondary information. Now, one of the simplest guarantees that we offer through amnesia is canonimity, and the idea here is that we transform its entry to a form where it becomes indistinguishable for other K minus one entry so if you have this very simplified medical database that you see here. In pseudonymized, there's a pseudonym, which is an arbitrary ID, and then there's this descriptive information that can act as quasi-identifier, which is the zip code, the age of the patient and their nationality, and then you have the diagnosis. Now, in this dataset, just by knowing some bits of this information, you can re-identify a person, right? If I know that Ivan is 20 years old, then just by looking here, I know that this is his record. There's only one who's 21 years old, so I know he suffers from heart disease. Now, the simplest way to anonymize this is to reduce the information to the information accuracy in the quasi-identifiers in such a way that groups of K, here K equals to four identical clusters appear in the anonymized data. So now if I know that Ivan is 20 years old and appears in this dataset, then I can only infer that his record is one of these first four, and I can no longer identify the exact record. So I can provide here a statistical guarantee that whatever you know as background information about a person, there will always be K-candidate records on the published dataset. So this is an irreversible transformation where whatever I know, I cannot really go back and discover the actual information, the actual record, and the actual information. The price I paid for this is that I have lost data accuracy. I have completely removed the nationality information, I have removed the list, the list of significant bits about the zip code. The zip code is actually a structured number where the list of significant bits, so more detailed administrative regions and as you remove them you go to broader administrative regions. And then I have removed exact age and I just have age categories. So I paid the price in terms of information accuracy and what I got was a guarantee that no one can be really identified. I will not get now in this webinar in the discussion about the strengths of different privacy guarantees. Even this guarantee can allow some information records and I will always be able to infer some things from the published data. The only way to not infer anything is not to publish anything, but this actually beats the purpose of data sharing, but I have reduced the probability of identifying someone in a formal way and with a strict DINT. So I can now treat this data statistically and not personal data that can be more widely used and shared. Now I will get to few technical details that I will need to then show you in the demo. I think the part that requires the most effort from the point of view of the user is how to guide the anonymization algorithm in reducing the accuracy of the descriptive information. I will go back and show you again here the example. Now the age of a patient who was transformed to an age category. What the algorithm would do is it will try to remove as little information as possible to achieve the desired privacy guarantee. So this will happen automatically but it needs a sim put from the user. A kind of generalization rules of replacement rules of specific values with more general values that makes sense. For example, the algorithm did not decide itself here to use these categories. At some point the user must give the algorithm age categories that make sense. Right, depending on the type of analysis, the domain, maybe there would be different age categories like less than 18, more than 18 or more than 65 depending if these age thresholds have legal extensions or they are important in the scientific domain field. So in some way someone must give directions to the algorithm on how to do these replacements. Also in the zip code, the algorithm needs to know that it should start by removing the least important bits. So this information is offered to the algorithm by what we call a generalization hierarchy. So the generalization hierarchy can be considered as a set of generalization rules where its specific value can be replaced with a more generic value in a way that makes sense. And also these rules are organized that we start from the bottom where the whole domain of different values appear and they are progressively reduced up to just one value. Having these rules organized like a tree allows the algorithm to guarantee that it will find a solution always that if it goes from the bottom to the root of the tree upwards, then always it will get a solution that has less values and greater chances of satisfying the privacy guarantee. So when preparing for the data minimization, the user has to create these rules for data replacement. I'm showing you here an example which is a geographic example that easily makes sense where it takes countries and groups them to some wider region and then to even wider regions. So Greece and Italy are parts of South Europe and South Europe along with North Europe make the whole Europe. Right, so this is simple information that needs to be given to amnesia and to actually any any, any other organization tools that relies on the generalization principle in order to produce meaningful results. I will not get into details there are ways that these different ways that can be used by the data minimization algorithm. For example, you can always go just one level up or in some branches you can go one level up and others not. And this makes it different if it's full domain generalization, like here, or you can have local recording where only some values are generalized and some others are not depending on if it's needed or not for then for getting the privacy guarantee that you want. The most complex, the more complex than an organization processes the better quality of the results, but sometimes it will take more effort on users, on users part or it will need more specialized algorithms for data analysis to be able to exploit the differences in granularity of the result. So, one thing that's that takes effort is creating these rules and after I'm done with the presentation I will show you a demo for me and so you have amnesia can help you to create them with less as less effort as possible. Now, apart from K anonymity, amnesia supports KM anonymity. KM anonymity is targeted to sparsely dimensional data. In the previous data example, you had two relational tables. These are common tables for databases that have structured information where each record has a fixed number of fields and every field is completed with information. But this is not the most common case in applications. For example, if you just have data from a retail store from a supermarket where different people went there and bought different types of products. Now I don't have exact products or I just have product categories to have a simple example. Then you get records that each record that refers to each person might have bought or interacted with just a few items from a very large domain right the supermarket has 20,000 products and you have bought like 20 or 30 different K anonymity relies on making these records identical and if you try to make identical the bills from a supermarket you actually use every kind of all useful information. So KM anonymity comes here and tells you you do not need to make identical but just guarantee that every combination of M different values appear K times in the whole data set. This means that if a third person knows and products that you bought from a supermarket, then it will always have K candidate records at the anonymized data so in this way it's a weaker form of anonymization but because the privacy guarantees not as strong as K anonymity but at the same time it allows you to make some meaningful anonymization of multi dimensional data which is not achievable else. So you can see an example here now a third person that knows that any of these facilities or manolis bought to specific products in the anonymized record, it will always find two records that have bought these two products, and this is the guarantee. These are the weaknesses. Again I will not go to the whole technical detail I think it goes beyond the scope of this webinar, but if you have any question, then we can discuss it in the Q&A later or just feel free to contact me through amnesia helpdesk or through my email and I can give you more detail on when one guarantees better suited than that. So this is how the method works for K anonymity, I will not get K anonymity, I will not get into that. Just to close with the limitations of amnesia, these are things that we know by design or through our experience from the helpdesk. So one big limitation that's true for amnesia and for other data anonymization tools is that users are still not familiar with anonymization techniques, so they do not know exactly what to expect from a tool and this is actually makes difficult in guiding them, because it's not just next next choices have to be made in different stages of the anonymization process, and someone has to be aware of what the steps are to to be able to apply the techniques effectively. Amnesia cannot decide itself on the private parameters, this question that comes quite often, what's the best value for K should I use this privacy guarantee or the other, these are not choices that can be taken automatically at least or even to have suggestions at least at this stage. My advice on this is to follow as a rule of thumb what statistical authorities do like your start or your local statistical authorities. They do not anonymize the data but they used to have rules about the granularity of information that is published. For example, they publish aggregating information for groups that have at least three or five entities inside them. So this could be a guide on how to set K for the anonymization process for anonymizing a specific data set. And the last thing is, at the moment we have K anonymity and KM anonymity and a lot differential privacy, it's different privacy guarantee, protects the data set in a different way and allows for different types of inferences to be made. Again, this decision has to be taken by the DPO and they need to understand what exactly is protected and what is not. So, mainly, amnesia limitations, highlight some limitations of anonymization and what can be automated and what cannot. So that was all the presentation and I think we have time so I will make a small demo on how amnesia works. I will stop sharing and share again. Fix the data set in a different way and allows for different types of inferences to be made. Again, this decision has to be taken by the DPO and they need to understand what exactly is protected and what is not. So, mainly, amnesia limitations, highlight some limitations of anonymization and what can be automated and what cannot. So that was all the presentation and I think we have time so I will make a small demo on how amnesia works. I will stop sharing and share again. Okay, this is amnesia, this is the first screen that you see when you open it and for this example, I'm sorry. Okay, I think it works now. So the first thing you need to do is load the data set. This is one of several amnesia works with several known data storage managers like Zenode and data server. So these are two options. It has a specific function for pseudonymizing daikon images which is under development. This is one of the newest tools that have not talked about it. So what we're going to do here is that we're going to load from local storage. And I have this comma delimited file. And here it has an import wizard, which is, you know, similar to, for example, Excel import wizard, where you just put the delimiter. Can you zoom in on all this? Can you zoom a little bit because we can't read or see clearly. If it is possible to zoom a little bit. Thank you. Yes, yes, yes, yes. Thank you so much. Okay, so you have these records that are delimited. I will use the simplest anonymization example which is canonimity, the other methods for disk based anonymization came anonymity and so on. But here we will treat this as a simple table. Now, these are simplified and synthetic medical records that came from some that resent UK medical records. The diagnosis codes that for which for this example we're going to treat them just as one value. Amnesia has the ability to treat them as different values. They actually set value which because they're an arbitrary number of different diagonal diagnosis. We have direct identifiers that will be unchecked and actually removed from the resulting data. And then we have the date of birth of patient and their market status. Now, just by removing these two fields. We can consider that the data should only mice, but we're going to go beyond that and actually anonymize the data so remove the fields. And we're ready to go on with anonymization. We do have some options about masking where but these are connected to should anonymization scenarios that I'm not detailing here but if you want we can just should minimize your data with Amnesia. Now, we're going to proceed to the most complex action that's needed for by the user which is the creation of these rules about generalizing values about replacing them with more generic ones. And we call this generalization hierarchy. Now, with Amnesia you can create the store and reload this kind of anonymization rules and I'm demonstrating here how these are depicted and what they are. And this was based on real data and they have the medical status and the patient or the doctor could write up any value brought forward by the patient. There's, there was no normal normalization of fixed value so somebody was, was importing know what the medical status. More information or less than needed. So, this was a way to group them to two different market status which was single and married and then this group to any. This is a very simple two step generalization hierarchy. Now, Amnesia can help you auto generate a hierarchy which works best for continuous values like eight salaries or even dates which is the most complex, the most complex continuous value that we have. And what Amnesia does it is it will read, you will choose. You will choose an attribute of the data that you loaded. I chose date of birth. It knows it's tie it produces a type of hierarchy and type of variable. It will calculate itself, the lowest and the biggest value that the minimum and the maximum values that exist in the data set. You need to give a name for this hierarchy. And then what will it would work like this. If you have values from a continuous domain. Think of a similar example first like the age of a person. Right. And that, you know, make its categories of that split ages in five year ranges. So you have zero to five five to 10 and so on. And then create a tree by grouping these states categories by two. So, original values will be replaced if needed by these ranges. Zero to five five to 10. And then the algorithm is creating more rules where it's category of the previous level will be grouped. It's two categories of the previous level will be grouped in one category of the next level so zero to five and five to 10 will be grouped together in zero to 10. And the next level zero to 10 and 10 to 20 will be grouped together on zero to 20 and so on till you cover the whole domain. And this creates a simple hierarchy where the algorithm that the algorithm can use to replace the original values. Now for dates this is more complex because we do not use the decimal system for dates right we have weeks that have seven days graph months that do not have exactly four or five weeks but something in between. And then we have years that have 12 months so we define the ranges in all three of these levels. We tell the algorithm to group days in groups of seven, then group months in groups of three, and then group the months in groups of one year. And from then on whether we use the decimal system about for years, group the years, group each category two categories to one at each level, and it creates a hierarchy like this. Now these are seven day periods where the exact dates will go here. The seven day periods are grouped to three month periods three months periods for three months periods a group to one year. And then we have to hear to hear periods for your periods eight year periods, 16 year periods and so on, till the whole domain is covered. And this has been auto generated just by the instructions that I gave to the algorithm. So we have created the hierarchies, we go to data and my station view and we instruct the algorithm to choose to anonymize the date of birth using the dates hierarchy that we just created to know to generalize the numbers with the marriage hierarchy that we loaded, and then to produce a two anonymous data set. Now, the algorithm will present us the whole solution space where the whole solution space is all different combinations of generalization that can happen to these two attributes the dates and the market status. This represents how many times each of these three attributes has been generalized. Here we know that the date of birth has been generalized three times it went to level three now generalization hierarchy, and the model status has been generalized to live twice. So we respond to solutions that provide us the the privacy guarantee that we wanted and read notes and solutions that do not provide the privacy guarantee that we asked the two anonymization. The reason that these are provided by the algorithm is that in some cases, let me see if this is one. Yes. We do not that despite having generalized three times the date of birth and once the model status, we cannot put all records in groups of two, but we're managed to put 99.8 of the records in groups of two, and it's only 0.2% of the records that do fall in these groups. So instead of further generalizing the data what we can do is just delete this 0.2% of the records and get an anonymous result by throwing out some outliers, but at the same time saving the quality of the rest of the records. And then you get the results like this, where the exact dates have been replaced by two year periods, the exact responses to marital status have been replaced by the first generalization of how to think this hierarchy the maritones or single, and you have a guarantee that it's a it's a record here it's combination of date of birth and marital status appears at least two times. So anyone seeing this record which this data set which is the anonymized data set will always have to candidate records for each person that he knows in advance. And let me see if we can find an example here. Well, if you go through all the pages of the data you will find some some of them marked with red, which will be the records that have been removed due to the last step of processing we took the separation. And this is all you save the data set and an organization show. So, this was all my presentation. I will be happy to respond to DNA now and you can also reach me through my email that you will see in the presentation and through amnesia help desk. Thank you for your attention. Thank you. Thank you so much, my knowledge for your presentation. And now I'll open the floor to all the participants if you have questions to address to the to my knowledge, feel free to do it. So, I have two questions but they are related more with with a service. How is the percentage of use of the office tool doing do you have and and and what what is the feedback from the users that you get when when they are using this amnesia. You get the first question the percentage of the views. What is the if you have a percentage of use of amnesia and what is the feedback you receive from the users. I have the numbers of use which were the totals were the ones I showed before the 110 k visitors and 400 k views at the last few years that it's up. The feedback we get we usually get the problems right in the help desk. It's great males I use it and it's great. I'm trying to use it and it doesn't work. So, and the difficult, many difficulties are on loading the data set and how you arrange and how you fix the values. And then there's some cases, usually not from the users that come to us come with some problem, which usually has to do with the first phase which is loading the data. And then it works for them, or maybe with loading the hierarchy or creating one, a few of them, but usually once they do this, they can anonymize the data and they do not come with the questions about anonymization itself. The questions about an organization itself they usually come from companies, we do get requests in amnesia help desk to present the tool and so they can see how they can use it. There are two kinds of here for companies they're consulting companies that come to us and they want to know what amnesia can do so they can integrate. Yeah, so use it part of a solution about privacy protection that they will offer to clients and tell them we can protect your data and for them and my says we can use this. And we have companies, big companies themselves that want to use it internally, and they come to us asking how they can they can use it and what exactly does in this presentation. We can presentations usually they need to understand that anonymization is the, it's not clear to them that you have to provide the statistical guarantee that's why I always in the presentation. I put a lot of effort in explaining what an organization is before explaining what amnesia does. So this is most of the feedback we get. I see also the question. At the moment we support just for simple data just CSV files, any kind of delimited, but CSV files. If you think as PSS data sets are important as a format for us, it's not difficult to add more formats but usually because for most data management programs it is easy to export CSV files we stay with with CSV files. I see next question about the limiting volume. The limitation, this algorithm that I showed you the most simple one keeps all data on main memory so the limitation comes from main memory it's not hard coded. If the data set is very big then you might get an error message. We have another algorithm, the disk based algorithm that has no limitation. If the data file feeds to the disk, then we can anonymize it so it can work for very big data. The most, I think the hardest constraint is on the size of the hierarchy. In some cases you can have hierarchies with millions of values. We do have, I did not show you in this scenario but if you use the ontologies from the internet as hierarchies for example hierarchy for the ICD the diagnosis codes. It can have thousands of values or you can have even larger ones with millions of values. There's no disk based method it has to fit in memory. If you have the hierarchy with millions of values which is not quite common, then you would need to have enough memory for your computer or for the Java, the Java virtual machine that answer. Thank you. I don't know if there's any more questions. If someone is using amnesia tool and have some questions, I think when all this is also sharing the help desk email help desk at open air at EU. I think you use the same service that open air is using and you receive all the questions through the help desk service. Yes. So I can leave here the email. I would be very happy if someone uses it to anonymize the data set to provide us some feedback if they're happy with the result if the result had the problem so if they wanted something different, something different. Even if it's not actually a problem of amnesia, please contact the help desk to let us know about the quality of the result or your experience from using it on a project. Yes. Thank you so much Manolis. I also left in the chat some links for the amnesia website for the guide and the fact sheet and amnesia video tutorials there are also available in YouTube. So you have your all the materials and the support materials that may help you and to try and test this powerful tool. So thank you so much again if there are no, oh, there's something thanking you in Portuguese. Thank you for the opportunity so if there is no more questions. Thank you so much again Manolis for being here and for your So detailed explanation and have a nice weekend and see you soon. Thank you for organizing this and thank you all for being here. Thank you.