A Review Of IDS-es In Large-scale IoT Systems: Challenges, Approaches, And Needs - Othmane Belarbi

uh hello everyone thank you for coming Welcome to our presentation so today I talked about the insurance Addiction in large scale iot systems so I'll mention the challenges approaches and needs so let's start written introduction so my name is I'm a first-year PhD student at Cardiff University and my research topic is about securing decentralized Federated learning based ideas with efficient local optimization so let's talk about the environment that I we would like to provide so is The Logical iot systems so there has been a rapid expansion of iot systems and it thanks to the iot devices ability to gather and communicate data which makes them very powerful The Cisco research are predicted in 2025 and 2030 38.6 billion and 50 billion iot devices connected to the internet and why I'm saying that is because most iot devices connected to the internet are quite exposed to cyber attacks due to weak security systems so even better this is the same Cisco research group has predicted that as the number of iot devices connected to the internet cases increases the number of cyber attacks will increase so that's why we need a system that will allow us to protect such system so an incident depiction system or ideas for short is a system that protects a network infrastructure or computers or malicious activities or any well-known threat so if others it follows the user or the network administrator Whenever there is good control suspicious activity in the network so here in that case the IDF is placed just after the firewall so imagine you have an incoming packet coming from the internet so before the packet arriving to the computer the ideas will check whether or not this package is malicious or as any suspicious A5 so traditional ideas based their operational machine learning on machine learning models trained centrally in the cloud and then distributed across multiple devices however the specific characteristic of large-scale iot systems brings new challenges that needs to be considered to deploy the ideas in an effective manner so for example in a large catering system we have a lot of devices that are highly connected so we are generating huge amount of data and most of the case most of the most of the time this data contains sensitive information so we need to be able to build an IDs let's go to respond in real time by training a model without clicking any sensitive information from the data collected from the sensors so there are different characteristic for example the fact that many iot devices has limited for a processing r or memory or battery so in that case the implementation needs to be likewise and suitable for that kind of device there are several ways to deploy that kind of architecture there is decentralized the distributed and decentralized decentralized architecture is the most common and famous One it consists in two phases first collecting the data from the clients and then second showing a chosen chosen machine learning or deep learning uh algorithm on the cloud on the on the cloud side this kind of deployment is usually done for small network with small scalability and it can simply explain because if the clients need to transmit their data to clocks to a cloud or any third party this transmission takes time and if you have the more clients you have the more transmission you will have to do the benefit and then increasing their both devices will lead to the server being the performer so it's definitely noticeable for our use case then there is the distribute architecture so this distributed based ID solution in which both the training and inference of the machine learning model is performed locally so it allows us to preserve the privacy of the data because we'll have one client we have one model per clients and however the main problem is that these models are not sharing any learning experience so it's limited to a per user experience and finally comes the Federated learning one so the written that I need is a new distributed uh learning Paradigm is used by Google years ago where multiple clients are using collaboratively to train a model without having any access to the raw data so it consists in three things so the first one the clients who train their local model based on their local data then the difference means no data but just the model updates the parameters of the model found after training and then this Cloud this software server will perform the aggregation to create a new Global model that is more effective so the particular CF this method is that since the rule data is not shared but just the modern days we are ensuring the data locality and the data privacy also it reduces the network overhead because we don't need to transmit beforehand any raw data with any third party and it's definitely suitable for a large network with high scalability so this method has attracted many interests in the application of mutual detection systems however it does come with its challenges so training in heterogeneous and massive Network which is challenges that is to be carefully considered and that would affect the performance of the model so recent Studies have shown that actually Federated learning is not that private even though we are not sharing any raw data it's impossible to leave some sensitive information from the mobile Case by performing some inverse operation on the gradients so the directed learning is uh is presenting some challenges that that and there is some there is still some work on it for example one potential solution for working with this leakage of information on the modern days will be to encrypt the communication Channel making impossible for the attackers to infer any sensitive information and finally so one work that I would like to work during my page it would be to work on the new version of iterative learning which consists of removing this Central entity so it would be decentralized operated learning and hopefully it could solve some issues like the scene open later because the central server is the entity orchestrating on the training so if we move it we are able to uh you know we're able to remove this single point of failure issue and also make applicable Federated learning in appear to parents as well so that's for my presentation thank you for instance [Applause]