Mémoire
| Résumé : | Today, it is not trivial to mix two of the great fields of computer science: artificial intelligence and cyber security. Intrusion detection systems can detect and identify attacks on a network. We have here a means of defense present in all network architectures requiring a minimum of security. But what happens when this defense is the target of attackers, is it robust enough not to distort its detection and let through attacks that should not, in any case, be able to infiltrate our network? This document focuses on the attacks related to the artificial intelligence used by intrusion detection systems that allows the analysis of the abnormal behavior that could have some messages exchanged in a network. Machine learning algorithms have not been designed with security standards in mind, their sole purpose is to analyze data provided as input and produce a result. This study will therefore highlight the different methods of attack and defense against adversarial attacks that threaten the proper functioning of our machine learning machines. A recent dataset on the benchmark will be used in our experiments to retrieve traffic data analyzed by an intrusion detection system in a computer network. In the last few years, a lot of articles about adversarial attacks and protection methods have appeared and we will understand this new attack pattern and evaluate the performances and risks it offers to hackers. This study will focus on Machine Learning since intrusion detection systems are often implemented with support vector machines. |
