Apprentissage fédéré pour la détection des intrusions

Abstract

With the rapid growth of information technology, especially the expansion of communication and social media, an ever-increasing number of devices are interconnected, this can lead to the disclosure of sensitive and private information. Therefore, anyone who uses a device associated with the internet is helpless in the face of the dangers presented by online hackers. Cybersecurity is becoming a daily struggle with the emergence of new attacks. Many devices find it difficult to schedule automatic updates ; they lack responsiveness measures like intrusion detection systems (IDS) to detect vulnerability Therefore, there is a need for a new intrusion detection approach that will provide security and can also prevent new intrusion attacks. My project is addressing this problem. Indeed, it relates to the strong implication in the use of modern techniques of artificial intelligence for the intrusion detection the classification of the network attacks. Thus, this report explains the achievements of a new alternative method to the detection of attacks through federated learning. The objectives to be achieved during this stage are research, dataset collection, data preparation, modeling and evaluation of federated models for network intrusion detection. We successfully implemented federated learning with two types of classifications. A binary classification using a federated convolutional neural network model on the IDS2017 dataset and two multi-class classifications with a federated multilayer perceptron on the IDS2018 dataset. For these two types of classifications, four strategies of the client’s choice were used. All clients are selected, clients are randomly selected, selection based on attack types or based on operating systems during a round. the results obtained for the binary classification with the four strategies worked well with an accuracy ranging from 87% up to 93% for the strategy where all customers are selected and depending on the types of attacks. For the multi-class classification the best result obtained is with the selection of all clients with an accuracy of 90%. This shows even in a real-world environment where machines crashing or power outages can occur, Federated Learning remains effective and robust in these situations.

Date	15 Feb 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Chamseddine Talhi (Supervisor)