Privacy enabled, Smart Contract driven Fair and transparent reward mechanism in Fede... | Devcon SEA

[Music] thank um thank you tatina for opening this because that's that's something that I want to carry forward what you just discussed um but before I do that first of all I mean it's a privilege and honor to be speaking at this conference and one of the most dynamic talented and perhaps vibrant ecosystem developer ecosystem so thank you thank you Devon for giving me this opportunity to speak about here um so my talk is based on um so my talk is based on one of the research work that me and my team had done earlier and and this paper received a number of citations both from Academia and uh you know industry combined so I thought maybe this is a good opportunity to share uh some of the thoughts that we had around it uh with this community and hopefully uh whatever we have started this community can take it forward and also you know um take it to completion because some of the things that we have started uh these are uh you know solving some of the complex challenges and I strongly believe and very much optimistic that the if this can be solved there will be this community so uh uh before I go further uh one of the key things that I wanted to highlight is I know there are a number of discussions and topics around decentralized AI you know pretty much decentralize everything uh but in this specific case uh I'm speaking about something slightly different uh it is about fed AI uh this concept uh is something that started few years back by Google where uh just giving you a quick primer on this that there are multiple multiple organizations uh can train uh the data uh sorry machine learning models locally and then they can share the models not the data and this is a very very important concept especially with the privacy and regulations where no institution want to share their respective data but they are at least willing to share uh the insights from that data set in form of a model and in this setup if you can see there are multiple organizations each train locally and then they share with the uh you know aggregator which combines it and hopefully the belief is that a combined model is much better performing than the uh you know single ones um how do we achieve it right um so one of the Sol proposed solutions that we have in the paper by the way a lot of the details is in the paper itself at this point I'm just highlighting a few of the key top points and also some of the challenges but one of the challenge with adoption of this kind of technology is everyone would like to know what's in for me and how do I get incentivized for something that I'm actually contributing to towards a global model so the the solution that we laid out is a smart contract based contribution analysis framework for FedEd learning then we also design and develop Tech techniques for feder contribution by the way this feder contribution term to the best of our knowledge did not exist before so we think we actually coined this term um and then we also have enhanced security for private messaging between the contributor if you saw the prior uh prior slide and the aggregator us using Dynamic key generation it uses traditional cryptography and finally leverage ethereum to build economics of onchain reward and penalty just in case some of the contributors are malicious um one of the key things I wanted to highlight and one before I went to this uh you know this is the stock Falls right in between intersection of blockchain and AI so some concepts are uh you know some of the slides you'll see it's purely machine learning and Qui quick highlighting the uh ml part of it an ml model at the underlying you know um at the pretty much Foundation layer is uh a set of metrics with set of numbers which you can actually uh you know um which are uh laid out in this format uh in case of Federated computation so we have a number of metrics if you can imagine in this in the diagram before we had a number of contributor giving their respective models and how do we come up with it one single number that identifies the contribution from each of the contributors so we use fited computation and if you have some of you are familiar with linear algebra it's about leveraging uh prous normalization that two takes two metrics and merges the two and comes out with one single number to give us the number final number finally how do we calculate it so within the aggregation layer after each round of aggregation Federated contribution is recorded on chain each of the clients receive that contribution the client goes back and does the evaluation again of that model against its own data set and sends back the performance the uh the actual computation performance back to the chain and then there's a smart contract that picks it up and does the comput does the analysis of that uh the the framework is fairly simple I mean again the all the details are in the paper however uh from a high level perspective there's one smart contract that manages the entire stack there are three different functions and one of the key functions in that is is uh how do you actually set the contribution which is right at the center of the uh you know slide okay I think I just five minutes any questions exactly some answers maybe any questions it's quite a fascinating topic about documenting the contribution any question from the audience yes we got one yes I get to spoll this amazing hi um so speaking of uh what you've done your research on what is um still open where um you maybe just name a few Fields within Federated AI where research still has to be done or are we done oh yes actually this is a very very active area of research um and some of the key topics that are being researched around identifying the malicious actors uh finding a way to incentivize what I just described is how do you incentivize the contributors to the global model um and also like there are different algorithms that are being discussed on how do you optimize the end to- endend Communications so broadly speaking there are a number of challenges at the same time there are a number of opportunities uh which are being explored at the moment we have a uh one that the guy in a white shirt yes hi thanks for the great presentation can you uh talk about some best practices on incentive design how to get clients to contribute checkpoints that are improvements and not going backwards on the original model uh that's a great question um in fact one of the main challenges in identify fing the right rewarding mechanism is the how do you actually assess the performance of the you know of the global model right so one of the challenge that we ran ran into uh was each client can have a different data set the quality of the data set could be different uh the size could be different uh the number of iterations could be different uh and so one of the potential practi is or one of the potential solution that we identified Again by the way just just for a disclaimer some of the work that we did uh was in a permission uh Network um and that uh normalizing some sort of global normalization of uh attributes the parameters and the training um algorithms uh could bring uh could be helpful in uh you know normalizing the the or maybe come up coming up with the the right metrics to combine otherwise you could see very varying performances which could be misleading um you know um could give you misleading results yeah we have time for one more question oh actually time's up all right thank you so much