Detecting Attacks on the Internet of Things Based on Multitasking Learning and Hybrid Sampling Methods

Abstract

The purpose of the study: To analyze and implement methods of multi-task learning and hybrid sampling of network traffic data to detect attacks in Internet of Things networks in order to improve the representation of minority classes and achieve data balance; compare the performance of various neural networks based on single-task and multi-task learning with hard and soft separation of parameters; implement weight optimization methods that provide automatic initialization and tuning of deep learning parameters for attack detection tasks in Internet of Things networks.Research methods: system analysis, modeling, deep machine learning.Results obtained: An approach to detecting attacks in Internet of Things networks based on multi-task learning is proposed. A comparison was made of the effectiveness of single-task learning models and multi-task learning models with hard and soft sharing of parameters. A hybrid sampling method is presented that combines random undersampling with oversampling based on a generative adversarial network. In addition, a weight initialization algorithm is implemented to eliminate imbalanced classification in IoT networks, ensuring high performance of the model for different classes of attacks represented in the dataset. Experiments were performed on different datasets, and the results showed that multi-task learning models outperform single-task learning for network traffic classification, achieving higher detection performance, especially for rare attacks.Scientific novelty: A new approach to detecting attacks in Internet of Things networks based on multi-task learning and hybrid sampling methods is proposed. An analysis and comparison of hard and soft parameter sharing in multi-task learning is carried out. The proposed approach aims to solve the problem of unbalanced traffic classification in IoT networks by random undersampling and synthetic sample generation using a pre-trained generative adversarial network model to achieve efficient data rebalancing.