Review on Generative Deep Learning Models and Datasets for Intrusion Detection Systems

Abstract

Intrusion detection systems (IDSs) play an essential role in defense of all networks and information systems around the world. IDS is one way of reducing malicious attacks. When attackers adjust their attack tactics and find alternative attack strategies, IDS must also develop through more advanced methods. Deep learning is a subfield of machine learning (ML) methods focused on learning results. A comprehensive review of various deep learning methods employed in IDSs is discussed first in this paper. Then a deep classification scheme is introduced, and the significant works recorded in the deep learning works are summarized. We performed a taxonomy survey of the deep architectures and algorithms accessible in these works and grouped such algorithms into three groups: hierarchical, composite, and generative. Afterward, a wide range of intrusion detection fields investigates selected deep learning applications. Finally, we address common types of datasets and frameworks.

Fault detection of a reciprocating plunger pump with fault-free data based on an unsupervised feature encoder and minimum covariance determinant

“…Unsupervised learning, which uses unlabeled data to learn the distribution of data or the relationship between data, is expected to solve the above fault detection problems. In recent years, researchers have carried out unsupervised feature extraction to achieve fault detection [7][8][9][10]. However, there are still some challenging issues to further improve fault detection performance.…”

Security of Federated Learning: Attacks, Defensive Mechanisms, and Challenges

“…In recent years, improvements in the implementation of Machine Learning (ML) models, have significantly increased the adoption of this technology in a wide range real-world systems that revolutionize almost all industries [1][2][3][4][5][6]. Despite ML's enormous success, many domains can only desire to benefit from it, but are unable to do so due to two significant obstacles: (1) concerns about clients' data privacy, as well as the laws and regulations that govern them, and (2) inability to develop a ML model because of insufficient data or high training overheads.…”