Federated transfer learning for intrusion detection system in industrial iot 4.0

• Published in: Multimedia tools and applications Vol. 83; no. 19; pp. 57913 - 57941
• Main Authors: N, Malathy, G, Shree Harish Kumar, R, Sriram, NR, Jebocen Immanuel Raj
• Format: Journal Article
• Language: English
• Published: New York Springer US 16.02.2024; Springer Nature B.V
• Subjects: Computer Communication Networks; Computer Science; Cybersecurity; Data Structures and Information Theory; Deep learning; Industrial applications; Industry 4.0; Internet of Things; Intrusion detection systems; Machine learning; Marine mammals; Multimedia Information Systems; Neural networks; Privacy; Semi-supervised learning; Servers; Special Purpose and Application-Based Systems; Track 6: Computer Vision for Multimedia Applications; Walrus optimization; Federated Learning; Semi-supervised learning; Deep Learning; Privacy of data; Federated Transfer learning (FTL)
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-18379-6

A major concern for Industry 4.0 is security issues because of several new cyber-security risks. In recent eras, various Deep Learning methods have been applied for intrusion Detection. On the other hand, these methods have to send the data to the centralized unit. This may alarm various problems associated with efficiency, privacy, and response duration. Furthermore, the data generated by the Internet of Things (IoT) gadgets are more and it is challenging to receive the labeled data as labeling the data is more time-consuming and expensive. This Masquerade several issues to the Deep Learning methods where labeled data is required. To prevail over these challenges a new mechanism has to be acquired. This paper proposes a new federated transfer semisupervised learning approach that takes both labeled and unlabelled data cooperatively. In the first phase, the data is preprocessed, and normalized, and optimal features are selected using Walrus optimization. In the second phase, to learn the representative and less dimensional features an autoencoder(AE) is trained on every gadget using private or local unlabeled data. Then the centralized cloud server aggregates the local model into a global autoencoder by applying Federated Transfer Learning (FTL). In the end, the cloud server consists of a semisupervised neural network with fully connected network layers to the global encoder which in turn trains the model with the readily available tagged data. Experiments were carried out with real-world industrial datasets and the proposed model ensures more privacy without sharing the local data, improved classification performance even with less tagged data, and less overhead in communication.