Un apprentissage fédéré avec une sélection de clients pour la détection d’intrusions réseau

Abstract

An ever increasing amount of sensitive data is produced and stored on personal machines. The centralized view that prevailed previously in machine learning is therefore changing and a strong emphasis is placed on privacy which can help researchers analyze and understand interesting phenomena and models. While the benefits of massive data processing are indisputable, they also represent a significant threat to the privacy of users who may suffer significant consequences. Our main objective in this research is to study users privacy problem and to propose solutions that preserve privacy. Traditional machine learning methods require that learning data be centralized in one place. One of the goals is to design, implement, and evaluate innovative federated and practical methods for private data analysis. Different sets of data, with different (sometimes complex) structures, will be taken into account. In this context, we will propose federated learning ; is a promising direction that uses a decentralized approach on a well-known dataset in the literature. To achieve the task, all participating entities, such as smartphone users, must share their datasets. Federated learning allows different entities to collaboratively learn a shared model without sharing their local data. We will study existing methods based on a deep Learning approach for network intrusion detection. Finally, we will offer a solution to the problem by proposing an optimization for our distributed learning algorithm based on two existing Learning paradigms, autoencoders and federated learning.We will detail our approach, its implementation by comparing its performance with existing methods.

Date	7 Feb 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Chamseddine Talhi (Supervisor)