 Next on the stage we have Akanksha, who is a senior data scientist from Red Hat. Thank you very much. Hello everyone. I believe most of us in this room have mobile phones, laptops, and use tons and tons of web applications, websites, so many things that are collecting your data with or without your consent. I mean, they are collecting this data to make our lives better, make our experience worthwhile while we are using these applications. But wouldn't it be better if we knew that this data is never going to be misused and still would give us awesome inferences and predictions while we are using all these applications? So that's what I'm going to do today. I'm going to introduce you to an amazing technique called homomorphic encryption, which can help ensure your privacy at all points by still making the most of the data science world. Before getting into this, I would like to take a moment to introduce myself. My name is Akanksha Duggal. I'm a senior data scientist in the emerging technologies group at Red Hat in the office of the CTO. I come from Boston and these are few places where you can contact me after my talk. To start with, since it's a hefty topic, I'll try to break it down into smaller pieces so that it's easier for everybody to understand. Multi-party computation is a cryptographic technique that allows multiple parties or even competitive parties each in possession of fragments of private data to participate in any computing and a specific result is computed using MPC-based algorithms. And what is homomorphic encryption? Let's start from encryption. Encryption was the process where we are scrambling the data so that only certain parties would have access to the real information that resides in the data. And homomorphic encryption would allow you to perform any computation on this encrypted data. It's a revolutionary concept that ensures data privacy and security at all points. So there are times where clients would not want to share their private data with the model owners and still would love to use the model to make inferences. Homomorphic encryption allows you to do everything as is by not seeing the actual data. Talking about where this comes in picture and who all can use it, there are tons and tons of new applications all around in almost every ecosystem where input data's privacy is the utmost concern. Starting from the data related to electric grids, transportation, healthcare, government, everywhere homomorphic encryption would be highly, highly required. And one of the most important use case in my opinion is in the healthcare sector HEPA laws are applied every time people do not follow privacy and security rules. To talk about what's the workflow that we are trying to follow here. So on the left side we have multiple data sources. So imagine if competitive companies were to train on one machine learning model at the same time ensuring that no other competitive party would know what their data looks like but still want to get the best results out of the model. This is the workflow that I personally think would work the best where we use Q-Flow pipelines to do the data pre-processing on each one of them's private data. Then once the data is encrypted, where the companies would know what the data looks like and Q-Flow pipelines would not have any idea, the notebooks would not have any idea. And after this process, we fit this all in the AI or the machine learning model whichever one we would like to put in here depending on the use case. And this is where homomorphic encryption would come in place and give out results without taking anybody's private information. So everything has advantages and disadvantages at the same time. Talking about advantages of homomorphic encryption, it allows you to not interact with the model owners or share your personal information at any point. You can still perform awesome inferences on the encrypted data. You can also have the data storage outsourced because there is low risk of data leakage and there's also reduced compliance costs such as EPA laws. You don't have to pay any of the costs if you are using the right encryption techniques. And talking about disadvantages, surprise, surprise, the computation is highly expensive. We tried to run homomorphic encryption on a very small data set and it was super expensive. I don't think this is highly applicable to the real-world use cases as of now unless you have unlimited resources. But one of the next steps in this project that me and my team are doing is a collaboration with Boston University where we are working with the hardware accelerators and we are trying to make this as seamless as possible so that we can also apply this to a bigger data set. I would stop right here and take questions and if you don't have a lot of time, I would love to chat with you after the talk as well. Thank you.