Training neural networks with captured real-world network data may fail to ascertain whether or not the network architecture is capable of learning the types of correlations expected to be present in real data.
In this paper we outline a statistical model aimed at assessing the learning capability of neural network-based intrusion detection system. We explore the possibility of using data from statistical simulations to ascertain that the network is capable of learning so called precursor patterns. These patterns seek to assess if the network can learn likely statistical properties, and detect when a given input does not have those properties and is anomalous.
We train a neural network using synthetic data and create several test datasets where the key statistical properties are altered. Based on our findings, the network is capable of detecting the anomalous data with high probability.
Samir Puuska, Tero Kokkonen, Petri Mutka, Janne Alatalo, Eppu Heilimo, Antti Mäkelä
Puuska S., Kokkonen T., Mutka P., Alatalo J., Heilimo E., Mäkelä A. (2020) Statistical Evaluation of Artificial Intelligence -Based Intrusion Detection System. In: Rocha Á., Adeli H., Reis L., Costanzo S., Orovic I., Moreira F. (eds) Trends and Innovations in Information Systems and Technologies. WorldCIST 2020. Advances in Intelligent Systems and Computing, vol 1160. Springer, Cham